Release 0.46.0

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Build the expected MCP output path with the platform separator
2026-10-19 23:47:30 +02:00 · 2026-10-17 14:31:59 -07:00 · 2026-10-17 14:27:35 -07:00 · 2026-10-17 14:13:04 -07:00 · 2026-10-17 12:55:01 -07:00 · 2026-10-17 12:30:46 -07:00
113 changed files with 3532 additions and 443 deletions
@@ -6,7 +6,7 @@
  },
  "metadata": {
    "description": "Agent skill and Claude Code plugin for the remove-ai-watermarks CLI.",
-    "version": "1.0.10"
+    "version": "1.0.12"
  },
  "plugins": [
    {
@@ -29,9 +29,9 @@ The unified maintainer entry point is `scripts/provider_oracles.py`; its workflo
 slot schema live in `docs/provider-oracles.md`. Multi-account and multi-network work
 selects each named slot explicitly from the gitignored repository-local
 `.oracle-slots.json` by default and snapshots it into a separate immutable batch
-manifest. A bounded provider check may continue through its finite set of authorized
-account slots after a refusal; do not add new logins, IP rotation, network-route
-failover, unbounded retries, or cross-session result merging. Never commit the local
+manifest. A provider check may use the finite set of configured account slots and surfaces
+after a refusal. Do not add new logins, IP rotation, network-route failover,
+unbounded retries, or cross-session result merging. Never commit the local
 slot file.
 Every manual result stays bound to the exact sanitized upload hash and preserves the
 provider response verbatim.
@@ -39,11 +39,14 @@ provider response verbatim.
 Provider surface selection is API-first. OpenAI and Microsoft use their official API
 adapters before their Web surfaces when the requested media and local configuration are
 supported. Google and Meta currently use Web because this tooling has no usable API route
-for them. Web remains an explicit second surface, never an automatic retry after an API
-refusal, quota response, transport error, or indeterminate verdict.
+for them. A request to use a provider covers every configured surface, including Web,
+without a second confirmation. An API refusal, quota response, transport error, or
+indeterminate verdict stays on that surface and never silently submits the file to Web.

 Google Web oracle automation uses only the user's existing authenticated real Chrome
-session and explicit `/u/N/` account index. OpenAI, Microsoft, and Meta Web automation
+session. Gemini uses an explicit `/u/N/` account index. The SynthID Detector at
+`https://synthid.com/` (`synthid-web`) uses the same Chrome profile and does not take a
+`/u/N/` index; its terms forbid driving the page as an automated API. OpenAI, Microsoft, and Meta Web automation
 uses isolated Playwright contexts; it must never attach to or copy state from the user's
 real browser. Proxy credentials are read only through a slot's named environment variable
 and never enter a manifest or log. Dotenv interpolation may derive routed proxy URLs from
@@ -79,9 +82,12 @@ The same rule applies to install hints: name the extra that actually makes the c

 Run `bash maintain.sh` from the repository root. The authoritative type gate is scoped to `src/`; full-project Pyright can exhaust Node memory on the ML dependency graph.

+The full gate outlasts the agent's 10-minute foreground shell cap, which kills it mid-suite. Run it detached under `nohup caffeinate -dimu` with its exit code written to a file and poll that file. To stop a stray run, kill its own PID, never `pkill -f pytest`, which also takes down other sessions' suites on the machine.
+
 The security scanner's exit status is authoritative, including a failure after a success
 message. `tests/test_maintenance.py` checks clean, vulnerable, and broken scanner
-runs through the real shell entry point. Recheck unresolved advisories using the
+runs through the real shell entry point, and pins the CI `security` job to the
+same scanner line with no `continue-on-error`. Recheck unresolved advisories using the
 procedure in `docs/development.md`; do not revive the old text-matching bypass.

 Boundary modules for cv2, Torch, and Diffusers may carry narrow per-file relaxations for unknown third-party types. Keep pure-logic files strict, preserve the local piexif stub, and fix real errors before widening a pragma.
@@ -1,6 +1,8 @@
 ---
 name: provider-oracles
 description: Run hash-bound Google, OpenAI, Microsoft, or Meta watermark-oracle checks for this repository. Use for provider verification, multi-account Gemini checks, isolated Playwright checks, or recording oracle evidence.
+metadata:
+  internal: true
 ---

 # Provider oracles
@@ -29,22 +31,24 @@ request permission for each account or file already covered by the task.
 Submit each frozen artifact at most once per slot in that pass. Preserve
 inconclusive results and disagreements between accounts. Stop when the bounded
 plan is complete or no authorized usable slot remains; report remaining rows.
-New logins, another provider, a new network route, or another pass after the plan
-is exhausted require authorization unless already requested. Account
+New logins, another provider, or a new network route require authorization.
+The oracles and slots already configured for the requested provider may be used
+again, including their Web surfaces, without a second confirmation. Account
 continuation does not authorize changing candidate pixels or erasing failures.

 ## Default surface priority

 Prefer an official API adapter whenever that provider, media type, and configured
-slot support it. Use Web only when no usable API route exists or the operator
-explicitly requests the Web surface as a second check. For OpenAI and Microsoft,
-the default plans are API then Web. Google and Meta currently have only usable
-Web routes in this tooling. Inspect the order with `plan PROVIDER`.
+slot support it. For OpenAI and Microsoft, the default plans are API then Web.
+Google and Meta currently have only usable Web routes in this tooling. Google's
+routes are `gemini-web` and `synthid-web`. Inspect the order with `plan PROVIDER`.
+The same request may use every configured surface.

 The priority is a selection policy, not automatic failover. An API refusal,
 quota response, transport error, or indeterminate verdict must be recorded and
-reported. Do not submit the same artifact to Web unless the operator separately
-authorizes that next surface.
+reported. The same request already covers every configured surface of that
+provider, including Web. Run Web as its own acknowledged check and keep its
+verdict separate. Do not replace the API result with the Web result.

 ## Route by surface

@@ -84,6 +88,27 @@ record that it is unavailable and continue with another authorized existing
 session. Ask the user only when no usable authorized session remains. Do not
 start a new login.

+### SynthID Detector
+
+Use the user's real authenticated Chrome. Do not use isolated Playwright: the
+portal's additional terms forbid using the page as an automated API, bulk
+verification, and quota evasion.
+
+1. Validate a `synthid-web` slot. It names `browser_profile` and `account_label`.
+   It does not use `google_account_index` or a proxy variable.
+2. Prepare an immutable batch outside the repository. Inspect images the same
+   way as for Gemini, and do not submit a file that still shows a provider logo.
+3. Open `https://synthid.com/` in the named Chrome profile. Accept the terms
+   dialog when it is shown. Do not start a new login. If the page requires
+   sign-in and that profile has no session, leave the row unsubmitted.
+4. Attach one prepared upload through the page's own file control.
+5. Read the settled label. `Watermark detected` is `detected`. `Watermark not
+   detected` is `not_detected`. Quota, sign-in, and unclear wording stay
+   `refused` or `indeterminate`.
+6. Record the verbatim response. The result covers partner SynthID from Google
+   DeepMind, Nvidia, OpenAI, and Kakao, and it does not name which partner. Do
+   not record it as a Gemini-only or OpenAI-only verdict.
+
 ### OpenAI, Microsoft, and Meta Web

 Use the isolated Playwright runner, never the user's real Chrome:
@@ -1,6 +1,8 @@
 ---
 name: signal-discovery
 description: Discover missing AI provenance and watermark coverage in the remove-ai-watermarks library from fresh or historical corpora, then validate detection and removal parity. Use for corpus gap mining, periodic signal audits, new-provider discovery, or deciding whether old retained data must be replayed. Do not use for ordinary per-file identification or removal.
+metadata:
+  internal: true
 ---

 # Signal discovery
@@ -0,0 +1,39 @@
+{
+  "name": "remove-ai-watermarks",
+  "version": "1.0.12",
+  "description": "Identify and remove AI provenance marks from images and video the user generated or edited, using the remove-ai-watermarks CLI.",
+  "author": {
+    "name": "wiltodelta",
+    "url": "https://github.com/wiltodelta"
+  },
+  "homepage": "https://raiw.cc",
+  "repository": "https://github.com/wiltodelta/remove-ai-watermarks",
+  "license": "Apache-2.0",
+  "keywords": [
+    "ai-watermark",
+    "watermark-removal",
+    "c2pa",
+    "synthid",
+    "provenance",
+    "gemini",
+    "sora"
+  ],
+  "skills": "./skills/",
+  "interface": {
+    "displayName": "Remove AI Watermarks",
+    "shortDescription": "Identify and remove AI watermarks and provenance metadata",
+    "longDescription": "Runs the open-source remove-ai-watermarks CLI locally to identify C2PA, IPTC, EXIF and TC260 provenance, visible AI-generator labels and supported invisible watermarks, then removes them from images and video the user generated or edited. It does not remove stock, agency or other third-party marks.",
+    "developerName": "wiltodelta",
+    "category": "Productivity",
+    "capabilities": ["Read", "Write"],
+    "websiteURL": "https://github.com/wiltodelta/remove-ai-watermarks",
+    "defaultPrompt": [
+      "Check this image for AI watermarks and provenance metadata",
+      "Remove the AI metadata from my generated video without re-encoding it",
+      "Remove the visible AI label from the corner of my image"
+    ],
+    "brandColor": "#A78BFA",
+    "composerIcon": "./assets/icon.svg",
+    "logo": "./assets/icon.svg"
+  }
+}
@@ -0,0 +1,15 @@
+# Local state and secrets never belong in a plugin bundle.
+.env
+.env.*
+!.env.example
+.oracle-slots.json
+.local-eval/
+.venv/
+__pycache__/
+*.pyc
+# Generated media and model artifacts.
+data/
+*.mp4
+*.onnx
+*.safetensors
+*.pt
@@ -62,7 +62,7 @@ jobs:
          PYTHON

      - name: Upload distributed release identity
-        uses: actions/upload-artifact@v7
+        uses: actions/upload-artifact@cf430e030ddbb5b0abf93d22962f4752f3646cd9 # v7.0.2
        with:
          name: distributed-release-${{ github.run_attempt }}
          path: release-identity.json
@@ -104,7 +104,7 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - name: Checkout the tap
-        uses: actions/checkout@v7
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
        with:
          repository: wiltodelta/homebrew-tap
          token: ${{ secrets.HOMEBREW_TAP_TOKEN }}
@@ -227,7 +227,7 @@ jobs:
    needs: resolve
    runs-on: ubuntu-latest
    steps:
-      - uses: actions/checkout@v7
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
      - name: Publish the agent skill when its version moved
        env:
          CLAWHUB_TOKEN: ${{ secrets.CLAWHUB_TOKEN }}
@@ -21,8 +21,8 @@ jobs:
    permissions:
      contents: read
    steps:
-      - uses: actions/checkout@v7
-      - uses: astral-sh/setup-uv@v7
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+      - uses: astral-sh/setup-uv@37802adc94f370d6bfd71619e3f0bf239e1f3b78 # v7.6.0
        with:
          python-version: "3.12"
      - name: Download derivative weights
@@ -36,9 +36,9 @@ jobs:
    permissions:
      contents: read
    steps:
-      - uses: actions/checkout@v7
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1

-      - uses: astral-sh/setup-uv@v7
+      - uses: astral-sh/setup-uv@37802adc94f370d6bfd71619e3f0bf239e1f3b78 # v7.6.0
        with:
          python-version: "3.12"

@@ -22,9 +22,9 @@ jobs:
    permissions:
      contents: read
    steps:
-      - uses: actions/checkout@v7
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1

-      - uses: astral-sh/setup-uv@v7
+      - uses: astral-sh/setup-uv@37802adc94f370d6bfd71619e3f0bf239e1f3b78 # v7.6.0
        with:
          python-version: "3.12"

@@ -15,9 +15,9 @@ jobs:
    runs-on: ubuntu-latest
    environment: pypi
    steps:
-      - uses: actions/checkout@v7
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1

-      - uses: astral-sh/setup-uv@v7
+      - uses: astral-sh/setup-uv@37802adc94f370d6bfd71619e3f0bf239e1f3b78 # v7.6.0

      - name: Verify release tag matches package version
        run: |
@@ -32,8 +32,8 @@ jobs:
  skill:
    runs-on: ubuntu-latest
    steps:
-      - uses: actions/checkout@v7
-      - uses: astral-sh/setup-uv@v7
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+      - uses: astral-sh/setup-uv@37802adc94f370d6bfd71619e3f0bf239e1f3b78 # v7.6.0
      - name: Sync dev environment
        run: uv sync --frozen --extra dev
      - name: Agent skill packaging tests
@@ -16,8 +16,8 @@ jobs:
  lint:
    runs-on: ubuntu-latest
    steps:
-      - uses: actions/checkout@v7
-      - uses: astral-sh/setup-uv@v7
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+      - uses: astral-sh/setup-uv@37802adc94f370d6bfd71619e3f0bf239e1f3b78 # v7.6.0
      - name: Sync dev environment
        run: uv sync --frozen --extra dev
      - name: Check C2PA soft-binding registry snapshot
@@ -29,6 +29,19 @@ jobs:
      - name: Ruff format check
        run: uv run ruff format --check

+  security:
+    name: lockfile vulnerability scan
+    runs-on: ubuntu-latest
+    timeout-minutes: 5
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+      - uses: astral-sh/setup-uv@37802adc94f370d6bfd71619e3f0bf239e1f3b78 # v7.6.0
+        with:
+          enable-cache: false
+      # Policy and triaged blocks: docs/development.md "Known security-gate blocks".
+      - name: Scan uv.lock with uv-secure
+        run: uvx uv-secure uv.lock
+
  test:
    name: test ${{ matrix.os }} py${{ matrix.python-version }}
    runs-on: ${{ matrix.os }}
@@ -56,8 +69,8 @@ jobs:
          - os: windows-latest
            python-version: "3.14"
    steps:
-      - uses: actions/checkout@v7
-      - uses: astral-sh/setup-uv@v7
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+      - uses: astral-sh/setup-uv@37802adc94f370d6bfd71619e3f0bf239e1f3b78 # v7.6.0
        with:
          python-version: ${{ matrix.python-version }}
      - name: Sync dev environment
@@ -69,8 +82,8 @@ jobs:
    name: versioned transport contracts py3.14
    runs-on: ubuntu-latest
    steps:
-      - uses: actions/checkout@v7
-      - uses: astral-sh/setup-uv@v7
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+      - uses: astral-sh/setup-uv@37802adc94f370d6bfd71619e3f0bf239e1f3b78 # v7.6.0
        with:
          python-version: "3.14"
      - name: Create clean runtime contract environment
@@ -91,8 +104,8 @@ jobs:
    runs-on: ubuntu-latest
    timeout-minutes: 5
    steps:
-      - uses: actions/checkout@v7
-      - uses: astral-sh/setup-uv@v7
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+      - uses: astral-sh/setup-uv@37802adc94f370d6bfd71619e3f0bf239e1f3b78 # v7.6.0
        with:
          python-version: "3.12"
      - name: Install ffmpeg
@@ -38,7 +38,7 @@ jobs:
    steps:
      - name: Download distributed release identity
        if: ${{ github.event_name == 'workflow_run' }}
-        uses: actions/download-artifact@v8
+        uses: actions/download-artifact@9000827ccba6bdab643e8b6fd33ac0654aef8333 # v8.0.2
        with:
          pattern: distributed-release-*
          path: distribution-identities
@@ -40,6 +40,8 @@ Before a release, read [`docs/release-and-distribution.md`](docs/release-and-dis

 The published Agent Skill is [`skills/remove-ai-watermarks/`](skills/remove-ai-watermarks/). Update it in the same change as any CLI edit that alters commands, extras, exit codes, mark keys, or the intended-use boundary. Install and publish steps: [`docs/agent-skill.md`](docs/agent-skill.md).

+The same skill tree is the Codex plugin (`.codex-plugin/plugin.json`). `remove-ai-watermarks mcp` is a standard-library MCP server that runs one CLI command per tool call; keep it dependency-free, because the MCP registry launches it as a bare `uvx remove-ai-watermarks mcp` that cannot add extras ([`docs/mcp-server.md`](docs/mcp-server.md)).
+
 ## Module architecture

 [`docs/module-internals.md`](docs/module-internals.md) is the canonical per-module map, including design decisions, thresholds, calibration history, incident records, and regression guards. Read the relevant section before changing a subsystem.
@@ -6,6 +6,15 @@ Remove AI provenance marks from images and video you generated yourself:
 - invisible pixel watermarks through direct local-format disruption or diffusion regeneration;
 - C2PA, EXIF, XMP, IPTC, and related AI metadata.

+| Before | After |
+| --- | --- |
+| ![Image with a visible Gemini watermark](demo_banana_before.png) | ![Image after visible watermark removal](demo_banana_after.png) |
+
+```bash
+uv tool install "remove-ai-watermarks[visible]"   # CLI for visible marks and metadata
+npx skills add wiltodelta/remove-ai-watermarks    # agent skill for Claude Code, Codex, Cursor, and other agents
+```
+
 Video support covers provenance identification, complete visible-plus-metadata
 cleaning, directory batches, visible Sora, Veo, Seedance, Doubao, Dola,
 Hailuo AI, and Kling AI mark removal, and oracle-certified VAE regeneration for video SynthID
@@ -274,11 +283,8 @@ features, and development setup.

 ### Visible Gemini mark

-| Before | After |
-| --- | --- |
-| ![Image with a visible Gemini watermark](demo_banana_before.png) | ![Image after visible watermark removal](demo_banana_after.png) |
-
-The `after` raster is generated from the tracked `before` raster by the public path:
+The `after` raster at the top of this page is generated from the tracked `before`
+raster by the public path:

 ```bash
 uv run remove-ai-watermarks visible demo_banana_before.png \
@@ -370,11 +376,11 @@ remove-ai-watermarks batch ./images --mode all
 Visible mark support includes:

 - Google Gemini and Nano Banana visible sparkle watermark;
- Doubao, the Jimeng wordmark and top-left `AI生成` pill, Qwen, Kling AI,
-  Yuanbao, Baidu, LiblibAI's bottom-center wordmark and compact top-left pill,
-  and RunningHub labels;
- one calibrated Microsoft top-right white AI-badge variant;
- one calibrated Samsung Galaxy AI label variant;
+- Doubao, the Jimeng wordmark and top-left `AI生成` layouts, Qwen, Kling AI,
+  Yuanbao, Dola, WorkBuddy, Baidu, LiblibAI's bottom-center wordmark and compact
+  top-left pill, and RunningHub labels;
+- calibrated Microsoft top-right white and beige AI-badge layouts;
+- calibrated Italian and Korean Samsung Galaxy AI disclosures;
 - a generic brand-less bottom-right `AI生成` fallback for OS/gallery-level
  AI-edit stamps that carry no vendor wordmark of their own (vivo and Xiaomi
  Gallery confirmed); synthetic-only calibration, weaker evidence than the
@@ -476,9 +482,23 @@ Claude Code can add this repository as a plugin marketplace:
 /plugin install remove-ai-watermarks@remove-ai-watermarks
 ```

+Codex reads the same skill through [`.codex-plugin/plugin.json`](.codex-plugin/plugin.json).
 The skill refuses stock-agency and other third-party paid-asset marks. See
 [the agent skill guide](docs/agent-skill.md).

+## MCP server
+
+`remove-ai-watermarks mcp` serves `identify`, `remove` and `batch` as Model
+Context Protocol tools over stdio, each a thin wrapper around one CLI command:
+
+```json
+{"mcpServers": {"remove-ai-watermarks": {"command": "uvx", "args": ["remove-ai-watermarks", "mcp"]}}}
+```
+
+See [the MCP server guide](docs/mcp-server.md).
+
+<!-- mcp-name: io.github.wiltodelta/remove-ai-watermarks -->
+
 ## ComfyUI

 The separate
@@ -490,6 +510,8 @@ invisible removal.

 - A missing local signal means unknown, not clean. Proprietary pixel
  watermarks may remain after metadata has been stripped.
+  [What each path leaves behind](docs/known-limitations.md#what-each-path-leaves-behind)
+  lists, per command, what may remain and how to check it.
 - Visible removal reconstructs a small region. Results depend on the background
  and selected fill backend.
 - Invisible removal changes the whole image and may alter faces, text, or fine
@@ -571,7 +593,8 @@ bash maintain.sh
 ```

 See [module internals](docs/module-internals.md) before changing a subsystem
-with documented invariants.
+with documented invariants. Report vulnerabilities as described in
+[SECURITY.md](SECURITY.md).

 ## License

@@ -0,0 +1,25 @@
+# Security policy
+
+## Supported versions
+
+Security fixes go into the latest release. Older releases are not patched;
+upgrade through the channel you installed from (PyPI or Homebrew).
+
+## Reporting a vulnerability
+
+Report vulnerabilities privately through
+[GitHub private vulnerability reporting](https://github.com/wiltodelta/remove-ai-watermarks/security/advisories/new).
+Do not open a public issue for an unpatched vulnerability. Include the affected
+version, the command or API call, and a minimal input that reproduces it.
+
+Vulnerabilities in a dependency, such as a CVE in a package from `uv.lock`, can
+be reported as a regular issue when the advisory is already public.
+
+## Dependency vulnerabilities
+
+Dependabot raises alerts and update pull requests for `uv.lock` and GitHub
+Actions, and CI scans `uv.lock` with
+[`uv-secure`](https://github.com/owenlamont/uv-secure) on every push and pull
+request. A vulnerable dependency is upgraded to a fixed release; advisories
+without one are tracked under
+[Known security-gate blocks](docs/development.md#known-security-gate-blocks).
@@ -0,0 +1,20 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" width="512" height="512" fill="none" stroke="url(#lg)" stroke-width="2" stroke-linecap="round" stroke-linejoin="round">
+  <!--
+    The raiw.cc brand mark: the Lucide "sparkle" and "x" glyphs.
+    Lucide, https://lucide.dev, ISC License, Copyright (c) Lucide Contributors.
+    Permission to use, copy, modify, and/or distribute this software for any purpose
+    with or without fee is hereby granted, provided that the above copyright notice
+    and this permission notice appear in all copies.
+  -->
+  <path d="M11.017 2.814a1 1 0 0 1 1.966 0l1.051 5.558a2 2 0 0 0 1.594 1.594l5.558 1.051a1 1 0 0 1 0 1.966l-5.558 1.051a2 2 0 0 0-1.594 1.594l-1.051 5.558a1 1 0 0 1-1.966 0l-1.051-5.558a2 2 0 0 0-1.594-1.594l-5.558-1.051a1 1 0 0 1 0-1.966l5.558-1.051a2 2 0 0 0 1.594-1.594z" fill="url(#lg)" fill-opacity="0.15"/>
+  <g transform="translate(1 17) scale(0.3)" stroke-width="5">
+    <path d="M18 6 6 18"/>
+    <path d="m6 6 12 12"/>
+  </g>
+  <defs>
+    <linearGradient id="lg" x1="0" y1="0" x2="24" y2="24">
+      <stop stop-color="#a78bfa"/>
+      <stop offset="1" stop-color="#06b6d4"/>
+    </linearGradient>
+  </defs>
+</svg>
@@ -6,9 +6,9 @@ qwen,Alibaba Qwen consumer app,visible text,data/fixtures/visible/qwen/provider-
 kling,Kuaishou Kling AI,visible text,data/fixtures/visible/kling/provider-original-direct.png,available,authenticated Kling IMAGE 3.0 original; Latin KlingAI 3.0 mark detected at confidence 0.43; TC260 producer maps to Kuaishou,subject to Kling AI terms; noncommercial fixture
 yuanbao,Tencent Yuanbao,visible two-line text,data/fixtures/visible/yuanbao/provider-original.png,available,authenticated Yuanbao original; detected by Yuanbao engine at confidence 0.49; TC260 producer metadata maps to Tencent,repository
 samsung,Samsung Galaxy AI,visible Italian text,data/calibration/samsung/samsung_gray_1.png,available,original provider capture and detector calibration,repository
-runninghub,RunningHub,visible text,,missing,no redistributable provider original found,
+runninghub,RunningHub,visible text,data/fixtures/visible/runninghub/example.png,partial,synthetic detector fixture available; no redistributable provider original found,repository
 baidu,Baidu,visible text and tag,data/fixtures/visible/baidu/provider-original.jpeg,available,detected by Baidu engine; distributed under the OpenNoMark repository's CC BY-NC-SA 4.0 license,CC-BY-NC-SA-4.0; excluded from Apache-2.0
-liblib,LiblibAI,visible wordmark,data/fixtures/visible/liblib/current-provider-output.png,partial,"authenticated current LiblibAI original carries a top-left AI生成 pill rather than the registered historical bottom-center LiblibAI wordmark; TC260 producer metadata maps to LiblibAI; historical positive remains missing",repository
+liblib,LiblibAI,visible wordmark,data/fixtures/visible/liblib/current-provider-output.png,partial,"authenticated current LiblibAI original carries a top-left AI生成 pill rather than the registered historical bottom-center LiblibAI wordmark; TC260 producer metadata maps to LiblibAI; a synthetic historical-wordmark fixture is available at data/fixtures/visible/liblib/example.png, but a historical provider positive remains missing",repository
 microsoft,Microsoft image output,visible top-right AI badge,data/fixtures/visible/microsoft/provider-original.png,available,authenticated Copilot original detected by Microsoft engine at confidence 0.54; also has valid C2PA and InvisMark,repository
 jimeng_pill,ByteDance Jimeng,visible AI pill,data/fixtures/visible/jimeng_pill/provider-published-example.jpg,available,"real Jimeng output embedded in page 29 of Zhonglun W&D's 2026 Bank Finance Legal Newsletter; direct jimeng_pill detection confidence 0.28; the same image also carries the Jimeng wordmark",fair-use research fixture; excluded from Apache-2.0
 google_synthid,Google,image SynthID,data/synthid/originals/Gemini_Generated_Image_633uuy633uuy633u.png,available,provider oracle recorded in data/synthid/manifest.csv,repository
@@ -1,8 +1,8 @@
 key,provider,watermark_type,path,status,verification,license
 sora,OpenAI Sora,visible moving mascot and wordmark,data/fixtures/visible/sora/provider-original.mp4,available,"Apache-2.0 upstream original; detected by Sora engine on 300/300 frames; C2PA claim names Sora and OpenAI/Truepic, with valid integrity and signature but an expired untrusted signer certificate",Apache-2.0
 veo,Google Veo,visible fixed text,data/fixtures/visible/veo/provider-original.mp4,available,"authenticated Google Flow Veo 3.1 Fast original; detected by Veo engine on 192/192 frames; valid C2PA and provider-confirmed SynthID",repository
-seedance,ByteDance Seedance,visible boxed AI label,,missing,"OpenRouter Seedance 2.0 Mini with watermark=true produced BytePlus C2PA but no visible boxed AI label; direct consumer original still needed",
-doubao,ByteDance Doubao,visible fixed text,,missing,no redistributable provider original found,
-dola,ByteDance Dola,visible fixed text,,missing,no redistributable provider original found,
+seedance,ByteDance Seedance,visible boxed AI label,data/fixtures/visible/seedance/example.mp4,partial,"synthetic detector fixture available; OpenRouter Seedance 2.0 Mini with watermark=true produced BytePlus C2PA but no visible boxed AI label; direct consumer original still needed",repository
+doubao,ByteDance Doubao,visible fixed text,data/fixtures/visible/doubao/example.mp4,partial,"synthetic detector fixture available; a real public carrier is retained locally but is not publication-cleared",repository
+dola,ByteDance Dola,visible fixed text,data/fixtures/visible/dola/example.mp4,partial,synthetic detector fixture available; no redistributable provider original found,repository
 hailuo,MiniMax Hailuo AI,visible composite label,data/fixtures/visible/hailuo/provider-original.mp4,available,"authenticated Hailuo AI MiniMax H3 original downloaded with With Watermark selected; detected by Hailuo engine on 124/124 frames; TC260 AIGC label names MiniMax",repository
-kling,Kuaishou Kling AI,visible fixed logo and wordmark,data/fixtures/visible/kling/provider-original.mp4,available,"authenticated Kling AI VIDEO 3.0 original; detected by Kling engine on 121/121 frames; no locally readable provenance metadata",Kling AI terms; noncommercial fixture
+kling,Kuaishou Kling AI,visible fixed logo and wordmark,data/fixtures/visible/kling/provider-original.mp4,available,"authenticated Kling AI VIDEO 3.0 original; detected by Kling engine on 121/121 frames; no container provenance metadata, but the Kling H.264 SEI tag (kling_sei) is readable",Kling AI terms; noncommercial fixture
@@ -26,6 +26,11 @@ SkillsMP indexes public GitHub `SKILL.md` files and does not need a separate
 upload. ClawHub and the Claude community marketplace are explicit publishes;
 see [Release and distribution](release-and-distribution.md#agent-skill).

+Every maintainer skill under `.claude/skills/` carries `metadata.internal: true`.
+The skills CLI discovers every `SKILL.md` in that directory, so without the flag
+`npx skills add` offers it and skills.sh lists it beside the published skill.
+`tests/test_agent_skill.py` fails on a project skill that lacks it.
+
 ## Where the skill can actually be listed

 Most skill catalogs are install registries, not discovery engines. A listing
@@ -65,6 +70,17 @@ Do these in order after the skill is on the default branch.
     are indexed via clawskills.sh, and the ClawHub prerequisite in step 6 is
     now met. The list still requires proven real-world usage. Re-attempt once
     the ClawHub listing shows organic installs.
+   - `hashgraph-online/awesome-codex-plugins`: READY, not submitted. It lists
+     only Codex plugins, so it needs `.codex-plugin/plugin.json` and the icon in
+     `assets/icon.svg`. Its contribution CI runs `scripts/validate-plugin-pr.py`
+     against the repository's `HEAD.zip`, then the HOL plugin scanner with a
+     minimum score of 80 and no high finding; that scan job is
+     `continue-on-error`, so it reports rather than blocks. The scanner's `\beval\s*\(`
+     rule also matches PyTorch's `.eval()`, which is why the code calls
+     `.train(False)`. On 2026-10-17 `plugin-scanner` 3.49.0 scored the tree 86
+     with no high finding. The remaining loss is a tracked video over the
+     scanner's 16 MiB file budget. Recheck with
+     `uvx --from plugin-scanner==3.49.0 plugin-scanner scan . --min-score 80 --fail-on-severity high`.
 5. **Claude community marketplace.** Form at
   [platform.claude.com/plugins/submit](https://platform.claude.com/plugins/submit)
   (individuals) or the claude.ai directory form (Team/Enterprise). Validate
@@ -186,8 +202,8 @@ that names a command, flag or value the previous release lacked. Left behind, it
 CLI is current right before it types an option that build rejects.

 Bump `metadata.version` in `SKILL.md` plus `version` in
-`skills/.claude-plugin/plugin.json` and `.claude-plugin/marketplace.json` when
-the skill instructions change. That version is independent of the PyPI package
+`skills/.claude-plugin/plugin.json`, `.claude-plugin/marketplace.json` and
+`.codex-plugin/plugin.json` when the skill instructions change. That version is independent of the PyPI package
 version.

 ## Behavior evals
@@ -1597,6 +1597,38 @@ Local reproducibility artifacts are under
 `contamination-scan-2026-08-26`, and `taxonomy-continuation-clean-2026-08-26`
 run directories.

+## Modern frozen-backbone campaign closed, 2026-09-27
+
+A second pre-registered representation campaign tested exact-revision,
+Apache-2.0 frozen vision backbones with universal ridge heads against the same
+product contract. The candidates were `google/siglip2-base-patch16-224` at
+revision `75de2d55ec2d0b4efc50b3e9ad70dba96a7b2fa2` and
+`facebook/PE-Core-B16-224` at revision
+`a16450b46fef32363459920c2685a1b4ef13dcd9`. The protocol was frozen before
+inference and used 28,743 unique source images. Its comparison cells contained
+1,873 AI-test images, 3,000 fresh Open Images photographs, 24 Kodak images, and
+300 FLUX images.
+
+| Model at its frozen operating point | AI test | Fresh photo false positives | Kodak false positives | FLUX |
+| --- | ---: | ---: | ---: | ---: |
+| Production Model 1 baseline | 1,676/1,873 (89.5%) | 35/3,000 (1.17%) | 0/24 | 256/300 (85.3%) |
+| SigLIP2 universal ridge | 1,602/1,873 (85.5%) | 74/3,000 (2.47%) | 0/24 | 125/300 (41.7%) |
+| PE-Core B universal ridge | 1,715/1,873 (91.6%) | 86/3,000 (2.87%) | 0/24 | 170/300 (56.7%) |
+
+SigLIP2 failed the AI-test, fresh-photo, and FLUX gates. PE-Core improved the
+AI-test cell but failed the fresh-photo and FLUX gates. Matching the baseline's
+exact 35-false-positive photo budget after the fact did not recover the lost
+coverage: SigLIP2 reached 1,505/1,873 AI and 98/300 FLUX, PE-Core universal
+reached 1,625/1,873 and 133/300, and PE-Core photo-only reached 1,709/1,873 and
+97/300. No threshold matched the baseline photo budget, AI coverage, and FLUX
+coverage simultaneously.
+
+The bounded newer-backbone hypothesis is therefore closed as a Model 1
+replacement. Keep `clip-l-ft.pt`, its ONNX graph, hashes, and Hub revision
+unchanged. Reproducibility artifacts are local under
+`.local-eval/synthid/ai-photo-2026-08-22/modern-representation-campaign-2026-09-27/`;
+the campaign did not enter the publication stage.
+
 ### Wild extras, not SynthID

 | Hypothesis | 2026-08-23 | Use |
@@ -28,6 +28,7 @@ defaults. This page focuses on choosing the right command.
 | `video invisible` and `video all --invisible` | `remove-ai-watermarks[video,diffusion]` plus ffmpeg on PATH |
 | HEIC/HEIF/AVIF pixel input | Add `remove-ai-watermarks[heif]` |
 | Every production command and Python-compatible backend | `remove-ai-watermarks[all]` |
+| `mcp` | Default package; each tool needs the extra of the command it runs |

 `batch` requires the same extra as its selected mode. Extras can be combined in
 one installation, for example `remove-ai-watermarks[visible,detect,heif]`.
@@ -469,6 +470,11 @@ detected. Use `--force` when you know the image should be processed:
 remove-ai-watermarks invisible image.png -o clean.png --force
 ```

+For transparent inputs, use PNG or WebP output to retain transparency. The engine
+preserves the original alpha channel, including partially transparent edges,
+while regenerating RGB pixels. JPEG output cannot preserve transparency. This
+also applies to `batch --mode invisible`.
+
 ### Choose a strength cohort

 `--vendor` selects the cohort the default strength resolves from. `auto`
@@ -685,6 +691,16 @@ For `visible` and `all` modes, the final summary includes counts for
 validation results are printed explicitly, but the once-filled outputs are kept
 and the batch does not retry them automatically.

+## Serve the CLI to agent clients
+
+```bash
+remove-ai-watermarks mcp
+```
+
+Runs a Model Context Protocol server on stdin and stdout whose `identify`,
+`remove` and `batch` tools each run one command of this CLI. Client setup and
+the tool contract are in the [MCP server guide](mcp-server.md).
+
 ## Exit behavior

 The CLI uses nonzero exit codes for meaningful incomplete outcomes, including
@@ -29,7 +29,7 @@ Do not add advisory ignores or suppress the scanner's exit status. Run and
 report Ruff, Pyright scoped to `src/`, and tests separately when security blocks
 the gate.

-Rechecked on 2026-09-10 against the locked versions and the advisory sources:
+Rechecked on 2026-10-02 against the locked versions and the advisory sources:

 - `lightning` and `pytorch-lightning`, pulled by the optional `trustmark`
  extra, are pinned to 2.6.6. That release clears PYSEC-2026-3624 and
@@ -37,18 +37,19 @@ Rechecked on 2026-09-10 against the locked versions and the advisory sources:
 - `accelerate`, required by the diffusion stack, is affected by
  [GHSA-4j2p-28q2-5m79](https://github.com/advisories/GHSA-4j2p-28q2-5m79)
  and PYSEC-2026-3804. The sharded-checkpoint loader accepts unsafe
-  `weight_map` paths; neither advisory lists a patched release. 1.15.0 is
-  on PyPI and is outside the advisory's `<= 1.14.0` range, but its
+  `weight_map` paths; neither advisory lists a patched release. The lock
+  carries 1.15.0, which is outside the advisory's `<= 1.14.0` range, so the
+  scanner no longer blocks the gate. Its
  `load_checkpoint_in_model` still joins `weight_map` entries with no
  containment check, and the sanitizer PRs
  ([#4138](https://github.com/huggingface/accelerate/pull/4138),
-  [#4214](https://github.com/huggingface/accelerate/pull/4214)) did not
-  land. Do not bump solely to silence the scanner. The project has no
+  [#4214](https://github.com/huggingface/accelerate/pull/4214)) were closed
+  without merging. Do not treat the passing scan as a fix. The project has no
  direct call to `load_checkpoint_in_model` or
  `load_checkpoint_and_dispatch`, but this does not prove indirect
  model-loading paths are unaffected.

-Before carrying the remaining block forward, rerun `uvx uv-secure uv.lock` and check
+When a new block appears, or before carrying one forward, rerun `uvx uv-secure uv.lock` and check
 PyPI for a fixed release. Upgrade a fixed package with
 `uv lock --upgrade-package <package>` and rerun the gate. Removing the current
 dependency would remove a supported optional feature, so that is not a
@@ -98,6 +99,11 @@ run the full-clip video test. Diffusion and model-running tests skip in that
 matrix; metadata, identification, visible removal, the DWT-DCT decoder, and the
 OpenCV eraser remain covered across operating systems.

+A separate `security` job runs the `maintain.sh` lockfile scan, so an advisory
+without a released fix turns `main` red; see
+[Known security-gate blocks](#known-security-gate-blocks). Vulnerability
+reporting is described in [`SECURITY.md`](../SECURITY.md).
+
 Keep `uv.lock` compatible with `uv sync --frozen`. Dependency pull-request checks use GitHub's merge result against current `main`; if `main` moves, merge it locally and rerun the full gate because a newer linter can expose stale directives in later code.

 `test.yml` runs one concurrency group per ref with `cancel-in-progress: true`, so a newer push cancels the previous run mid-flight. A cancelled run is not a failure verdict: the superseding run's head includes the cancelled commits as ancestors, so watch that run to completion instead of retrying the push.
@@ -27,7 +27,8 @@ to run the tool. Use the maintainer references only when changing the code.
 | [Code provenance](code-provenance.md) | Required notices for licensed derivative work. |
 | [Verification plan](verification-plan.md) | Verification methods, completed measurements, and remaining validation gaps. |
 | [Release and distribution](release-and-distribution.md) | PyPI, Homebrew, Hugging Face Space, photo-classify Hub model, ComfyUI Registry, and the release workflow. |
-| [Agent skill](agent-skill.md) | Install and maintain the published Agent Skill and Claude plugin. |
+| [Agent skill](agent-skill.md) | Install and maintain the published Agent Skill, the Claude plugin and the Codex plugin. |
+| [MCP server](mcp-server.md) | Connect an MCP client to `remove-ai-watermarks mcp`, or publish it to the MCP registry. |
 | [Watermarking landscape](watermarking-landscape.md) | Vendor signals and detection approaches. |
 | [SynthID technical reference](synthid.md) | Mechanism, provenance, robustness, regeneration. |

@@ -4,6 +4,21 @@ This page describes current product limits. Historical measurements and
 superseded experiments live in the research archive listed in
 [the documentation index](index.md).

+## What each path leaves behind
+
+No command produces a "clean" verdict. Each one removes a specific layer, and
+the remaining layers need their own check.
+
+| Path | Removes | May remain | How to check what remains |
+| --- | --- | --- | --- |
+| `metadata --remove`, `video metadata --remove` | Supported AI metadata: C2PA manifests, TC260 labels, IPTC AI fields, app-export JSON, generation parameters | Every pixel and audio watermark (SynthID, Meta Content Seal, InvisMark, TrustMark, DWT-DCT). A C2PA soft binding can still lead a resolver to a manifest stored remotely. Strings inside encoded video streams, such as H.264 SEI messages, stay too, except Kling's SEI tag, which `video metadata --remove` blanks | `identify` re-reads the metadata. Only the provider's verifier can confirm a pixel watermark. The pixels are not regenerated, so an AI-image classifier scores the output the same as the input |
+| `visible`, `erase`, `video visible` | The selected label region, refilled. AI metadata too, unless `--keep-metadata` is given | Any invisible watermark across the rest of the frame. The fill can leave a visible patch on structured backgrounds | Look at the output. Use `identify` for metadata and the provider's verifier for invisible marks |
+| `invisible`, `all`, `video invisible` | Regenerates pixels to disrupt invisible watermarks. `all` also removes visible marks and metadata | No local decoder confirms that SynthID is gone from a given file, and the outcome depends on content and seed. `video invisible` copies the audio track unchanged, so an audio watermark survives | The provider's verifier, such as Google's SynthID Detector or OpenAI's provenance check. A pass there says nothing about AI-image classifiers |
+| `identify`, `video identify` | Nothing | Not applicable | `unknown` means no supported local signal was found. It is not a clean verdict |
+
+AI-image classifiers are a separate question from provenance. They score how
+the pixels look rather than read a mark, and none of these paths targets them.
+
 ## Pixel classification

 `classify` / `classify_pixels` is AI versus camera, not a universal AI detector.
@@ -179,9 +194,12 @@ fresh-font and silhouette variants visibly changed typography. The higher-fideli
 and line geometry. It builds its donor internally but still requires an independently
 clean global anchor. Automatic OCR and line-box proposals are not reliable enough to
 remove those requirements, and exact oracle results do not establish a general mask,
-seed, runtime, or provider operating range. Qwen-Image-2.0 is hosted-only and exposes
-no equivalent low-strength denoise control. Exact experiments, controls, and pass
-rates are kept in
+seed, runtime, or provider operating range. Qwen-Image-2.0 is hosted-only. Qwen-Image-2.1
+(issue #111) has open weights, but its Diffusers pipeline uses `image=` only as
+reference conditioning and starts the target latents from noise, with no `strength`
+or latent noise-mask input. Its Qwen Research License also limits the weights to
+non-commercial research and evaluation, so it is not a supported-profile candidate.
+Exact experiments, controls, and pass rates are kept in
 [`text-protection-research.md`](text-protection-research.md) and the
 [`fidelity` evaluation record](../data/evaluations/fidelity/README.md).

@@ -260,7 +278,9 @@ guarantee, and check the audio separately when it matters. The question is no
 longer Google-only: since 2026 OpenAI (ChatGPT and API audio) and ElevenLabs
 also embed SynthID in generated audio, and Microsoft lists provenance for its
 Azure speech and GPT audio models, so voice-over or music from those services
-muxed into any video survives this path unchanged.
+muxed into any video survives this path unchanged. Suno announced audio
+watermarking and fingerprinting on 2026-08-06 without naming the technology, so
+Suno music belongs on the same list.

 The public video results make that boundary explicit without decoding audio:
 `visual_invisible_action` records whether video pixels were regenerated, while
@@ -397,12 +417,20 @@ global Qwen stage is tiled; the face stage runs after tile blending.
 every profile, streaming weights instead of pinning them. It reduces CUDA memory
 pressure at the cost of speed.

-It reaches the global stack of `qwen-zimage` only. That stack is each
-profile's own, and `chroma-zimage` and `sdxl-zimage` load theirs with a plain
-`.to(device)`. On a card below the face-stage residency floor, where the face
-stack was already offloading, the flag changes nothing at all for those two: the
-run behaves as if it were absent. The library says so on the way past, as a
-warning ahead of the model load.
+It reaches the global stack of `qwen-zimage` and `sdxl-zimage`. That stack is
+each profile's own, and `chroma-zimage` loads its stack with a plain `.to(device)`.
+On a card below the face-stage residency floor, where the face stack was already
+offloading, the flag changes nothing at all for that profile: the run behaves as
+if it were absent. The library says so on the way past, as a warning ahead of the
+model load.
+
+`sdxl-zimage` streams its global stack without being asked on a card below
+`RESIDENT_SDXL_GLOBAL_MIN_VRAM_GIB` (12 GiB). Its UNet and Canny ControlNet run
+together on every step, so the stack does not fit an 8 GiB card even with
+model-level offload. A resident load there does not always fail loudly: under
+the Windows display driver, including WSL2, CUDA can spill the overflow into
+shared system memory, and the run goes on at a fraction of the speed instead of
+raising an out-of-memory error.

 Residency is otherwise chosen from the card's total VRAM. On a card large enough
 to hold a stack, offloading is pure waste: DiffSynth drops weights to the meta
@@ -0,0 +1,91 @@
+# MCP server
+
+`remove-ai-watermarks mcp` runs a [Model Context Protocol](https://modelcontextprotocol.io/)
+server on stdin and stdout, so an agent client can identify and clean the
+user's own images and video without a shell. It is a thin wrapper: each tool
+runs one command of this CLI in a child process and returns that command's exit
+code, output path and text.
+
+The subcommand ships since 0.46.0. The server uses only the Python standard
+library, so the default package runs it. The commands behind a tool need the
+same extras they need on the command line
+([CLI guide](cli.md#command-dependency-map)). A missing extra comes back as a
+tool error carrying the CLI's install hint.
+
+## Client configuration
+
+Most clients take a command and arguments. With [uv](https://docs.astral.sh/uv/):
+
+```json
+{
+  "mcpServers": {
+    "remove-ai-watermarks": {
+      "command": "uvx",
+      "args": ["remove-ai-watermarks", "mcp"]
+    }
+  }
+}
+```
+
+To make visible-mark removal and pixel detectors available, name the extras:
+
+```json
+{
+  "mcpServers": {
+    "remove-ai-watermarks": {
+      "command": "uvx",
+      "args": ["--from", "remove-ai-watermarks[visible,video]", "remove-ai-watermarks", "mcp"]
+    }
+  }
+}
+```
+
+An existing installation can be started directly as `remove-ai-watermarks mcp`.
+Paths in tool calls are read and written on the machine that runs the server.
+
+## Tools
+
+| Tool | Runs | Writes |
+| --- | --- | --- |
+| `identify` | `identify --json`, or `video identify --json` for a video suffix; `check_visible: false` adds `--no-visible` | Nothing |
+| `remove` | `metadata --remove`, `visible` or `all` for an image; `video metadata --remove`, `video visible` or `video all` for a video | One output file, `<name>_clean` beside the source unless `output` is given |
+| `batch` | `batch --mode` for images or `video batch --mode` for video | A separate output directory, `<dir>_clean` unless `output_dir` is given |
+
+`mode` is `metadata` (the default), `visible` or `all`. For an image, `all`
+adds the invisible-watermark scrub where the installed extras and a CUDA GPU
+allow it; `video all` removes visible marks and metadata only. The server never
+writes over an input: an `output` equal to the source, or an `output_dir` equal
+to the input directory, is refused before anything runs.
+
+Exit codes keep their CLI meaning. `0` and `2` are results, not tool errors, and
+`2` means no supported signal was found, which is unknown rather than clean.
+Click also exits `2` on a usage error such as a missing file; its `Usage:` line on
+stderr marks that case as a tool error. Every other nonzero exit code is returned
+with `isError: true` and the command's stderr.
+
+## Scope
+
+The server carries the project's intended-use boundary in its `instructions`:
+it is for content the user generated or edited themselves, and not for stock,
+agency, marketplace or other marks that protect a third party's asset. See
+[legal and safety](legal-and-safety.md).
+
+## Protocol
+
+The server answers the initialize handshake for protocol versions 2025-11-25,
+2025-06-18, 2025-03-26 and 2024-11-05, and lists its tools with input schemas
+and annotations: `identify` is read-only, and `remove` and `batch` write new
+files without destroying inputs. It does not implement `server/discover`. A
+client that probes it first, such as the official Python SDK's `Client` in its
+default mode, receives method-not-found and falls back to the handshake.
+`tests/test_mcp_server.py` runs that client against the real subcommand.
+
+## Registry listing
+
+[`server.json`](../server.json) describes the PyPI package for the
+[official MCP registry](https://registry.modelcontextprotocol.io/) under
+`io.github.wiltodelta/remove-ai-watermarks`. The registry verifies ownership by
+the `mcp-name:` marker in `README.md`, which becomes the PyPI description, so it
+can only verify a release whose README carries that marker.
+[`glama.json`](../glama.json) names the maintainer for Glama. Publishing steps
+are in [release and distribution](release-and-distribution.md#mcp-registry).
@@ -47,6 +47,9 @@ Important contracts:
  `EXIT_NO_VISIBLE_MARK`.
 - `invisible` writes no output when no supported local signal is found, unless
  `--force` is supplied.
+  It decodes the source before the scrub gate and exits `1` when it cannot,
+  because the gate reads undecodable bytes as "no target" and would otherwise
+  report the no-signal code for a file it never read (fixed 2026-10-17).
 - The two no-signal conditions currently share exit code `2`.
 - A quiet `metadata --check` and a successful `metadata --remove` (same for
  `video metadata`) end by repeating the `identify` limit: the pixel channel is
@@ -236,10 +239,46 @@ payloads; TC260 removal changes the four-byte key to `free` and blanks only the
 validated JSON value with same-length spaces. Keyed generation metadata such
 as `workflow` and `prompt` uses the same structural walk; removal blanks the
 complete key and value in place. This preserves every box size,
-`stco`/`co64` offset, encoded stream byte, and source-sized memory bound.
+`stco`/`co64` offset, encoded stream byte apart from the Kling SEI tag below, and
+source-sized memory bound.
 Publication is atomic, and a malformed top-level walk is copied unchanged. A
 generic `AIGC` key whose value has no TC260 field is ignored.

+App-export JSON in a keyed or keyless metadata item is blanked the same way.
+Dreamina's MP4 export writes it twice, under `LvMetaInfo` and
+`com.apple.quicktime.artwork`. The file-level scan read both from the raw head
+bytes while the stripper walked only TC260 and generation keys, so
+`video metadata --remove` exited 1 with both markers still present on the
+tracked Dreamina fixture (found 2026-10-17). Its `UserComment`
+(`{"aigc_label_type":0,"source_info":"dreamina"}`) and `comment` (`vid:...`)
+items carry no marker the scan counts, so they stay.
+
+`test_no_flagged_metadata_value_survives_in_the_output_bytes` strips every
+tracked MP4 fixture and searches the output byte for byte for each flagged
+source value, after the pattern of byte-level leftover checks in public video
+scrubbers. The walk reaches container metadata only. Strings inside the encoded
+streams are out of its reach with one exception: x264 writes its version and
+options as an H.264 SEI message, which is not AI provenance and survives the
+stream copy; only a path that re-encodes the video stream (`video visible`,
+`video invisible`) replaces it.
+
+The exception is Kling's encoder tag. Every Kling video in the tracked fixtures
+and local captures (Kling 2.6, 3.0, 3.0 Omni and O3, through Higgsfield, Runway
+and Kling itself) carries one H.264 `user_data_unregistered` SEI message, payload
+type 5, whose 25-byte payload is UUID `91ca6061-4aee-3854-8614-2d5f73f4ae2e`,
+`kling-ai` and a NUL. `get_ai_metadata` reports it as `kling_sei` from the head
+window, and `identify_video` names Kling from it when no TC260 label does, as on
+`data/fixtures/visible/kling/provider-original.mp4`, which carries no other
+signal. The streaming stripper finds every copy of the UUID plus `kling-ai` in
+each top-level `mdat` during the copy pass it already makes, so the file is still
+read once, and overwrites those 24 bytes with spaces. The NAL length,
+SEI header and payload size stay, and a run of 0x20 cannot form an
+emulation-prevention sequence, so the message stays well-formed and no decoder
+reads it: `ffmpeg -f framemd5` hashes of every frame match before and after on
+both tracked Kling fixtures (found and fixed 2026-10-17). The QuickTime `©cmt` item in
+the Grok fixture (`Signature: <base64>`) also survives, because the xAI
+signature rule requires the UUID half of the pair and that half is absent here.
+
 [`_internal/ebml.py`](../src/remove_ai_watermarks/_internal/ebml.py) provides the
 corresponding bounded Matroska/WebM reader. It seeks over clusters and accepts
 only a `Segment.Tags.Tag.SimpleTag` pairing `TagName=AIGC` with a JSON
@@ -637,6 +676,10 @@ the official C2PA registry. Refresh it with
 upstream schema invariants, records the exact source revision and writes a
 deterministic module. Runtime inspection remains offline, and registry membership
 is name-only evidence rather than proof that a compatible decoder ships.
+`--check`, run by `maintain.sh` and the CI lint job ahead of Ruff, compares the
+snapshot with the live upstream revision, so an upstream registry commit fails
+lint on `main` and on every pull request until the snapshot is refreshed, and the
+Ruff step behind it does not run until then.
 The GitHub API revision lookup sends `GITHUB_TOKEN` (or `GH_TOKEN`) when set,
 because the unauthenticated 60 requests/hour per IP limit is exhausted on
 shared CI runners; `.env.example` documents the variable.
@@ -743,8 +786,11 @@ Key contracts:
  A.1 string (`"Label":"1","ContentProducer":"...","ProduceID":"..."`) under the
  default SM2 user ID, and MiniMax's `LabelMataData` spelling is accepted beside
  the guide's `Md`. Only the label signature is checked: MiniMax's binding
-  signature also verifies, but no verdict reads it, and its content hash is not
-  recomputed because MiniMax names no `CntSel`. The key travels in the file, so
+  signature also verifies, but no verdict reads it. The guide's appendix A.2
+  defines `DFT` for non-JPEG/PNG files as all bytes except the metadata label;
+  MiniMax names no `CntSel`, and measured whole-file, direct DFT byte-range,
+  metadata-box, and media-payload interpretations do not reproduce its content
+  hash. The key travels in the file, so
  only a key pinned in `TC260_PRODUCERS` (the source of `TC260_SIGNING_KEYS`)
  names a signer, and a pinned signer
  is checked with the pinned point whatever parity the file claims; an unpinned
@@ -1498,6 +1544,15 @@ schemas 0-2 count as positives. Schema 3 is below the precision threshold: all
 same false payload after re-encoding. The official Adobe Variant P schema-1
 fixture in `data/fixtures/provenance/` is the positive regression control.

+An independent benchmark supports gating beyond a valid codeword. Kazempourradi
+et al. ([arXiv:2610.04151](https://arxiv.org/abs/2610.04151), section 4.4)
+found that the blind detectors of TrustMark B, C, P and Q in BCH_5 mode reported
+a valid codeword on at least one of four decodes for 5.9% (P) to 15.3% (C) of 85
+unmarked images, and on 3 of 8 images in a 16-bit HDR subset. Testing the same
+decodes against the expected payload produced no false match. This detector is
+blind and has no expected payload, so the re-encode stability and schema
+restriction above stand in for that test.
+
 #### Why the DWT-DCT decoder looks the way it does

 Two things in `dwt_dct.py` are load-bearing and neither is obvious from the code.
@@ -1593,8 +1648,8 @@ their own: the Gemini sparkle has its own higher-confidence path, and the pill
 alone is too weak to attribute.

 The set of marks that veto the Jimeng pill is derived from the registry rows:
-every other product from the same manufacturer. Doubao therefore vetoes the pill
-because both products are ByteDance, while Qwen, Kling, Yuanbao, RunningHub,
+every other product from the same manufacturer. Doubao and Dola therefore veto the pill
+because all three products are ByteDance, while Qwen, Kling, Yuanbao, RunningHub,
 Baidu, and LiblibAI do not become siblings merely because they use the same TC260
 standard. Marks from other manufacturers cannot enable or veto the
 ByteDance-specific arm.
@@ -1653,8 +1708,7 @@ assets and template ladder are immutable.

 A decorative sparkle/glint/star glyph in stylized typography -- a bright,
 near-white four-point accent, the same silhouette family the engine's own
-template captures -- is a known, unresolved false-positive class. Incident
-2026-08-27 (raiw-app, Spaces uid 7e01ef8cb99949bba14ab4607a66ae06): a book-cover
+template captures -- is a known, unresolved false-positive class. A book-cover
 title's decorative diamond accent inside a neon-glow letterform scored 0.64, just
 under `_SPARKLE_FP_CONF`, and the removal step painted a permanent smudge over
 the artwork. Investigation (margin, gradient, saturation, and a prototyped local
@@ -1694,6 +1748,11 @@ Regression coverage:
 [`_text_mark_engine.py`](../src/remove_ai_watermarks/_text_mark_engine.py)
 provides common localization, detection front ends, template caching, rival
 comparison, and footprint construction.
+[`_text_mark_variants.py`](../src/remove_ai_watermarks/_text_mark_variants.py)
+combines independently gated layouts under one registry key while retaining the
+winning layout's template and mask geometry. Its bottom-right variant searches
+only anchor-valid placements, so an unrelated stronger match elsewhere in the
+corner ROI cannot hide the real edge-hugging mark.

 Each vendor module supplies a `TextMarkConfig` and only the behavior that cannot
 be represented by the shared base:
@@ -1703,6 +1762,8 @@ be represented by the shared base:
 - [`qwen_engine.py`](../src/remove_ai_watermarks/qwen_engine.py)
 - [`kling_engine.py`](../src/remove_ai_watermarks/kling_engine.py)
 - [`yuanbao_engine.py`](../src/remove_ai_watermarks/yuanbao_engine.py)
+- [`dola_engine.py`](../src/remove_ai_watermarks/dola_engine.py)
+- [`workbuddy_engine.py`](../src/remove_ai_watermarks/workbuddy_engine.py)
 - [`samsung_engine.py`](../src/remove_ai_watermarks/samsung_engine.py)
 - [`runninghub_engine.py`](../src/remove_ai_watermarks/runninghub_engine.py)
 - [`baidu_engine.py`](../src/remove_ai_watermarks/baidu_engine.py)
@@ -1747,7 +1808,10 @@ on 2026-09-04 uses `KlingAI 3.0`. The Latin template scored 0.429 on that 1024 x
 across the 94 available neighboring real and synthetic image controls, while an
 adversarial solid corner blob reached 0.379, so its strict gate is 0.40. This is
 one-positive provisional calibration, intentionally separate from the older
-CJK detector rather than presented as broad recall evidence.
+CJK detector rather than presented as broad recall evidence. Both layouts now
+project row-wide and column-wide structure out of the winning patch before a
+second correlation. This preserves glyph detail while rejecting ordinary panel
+and image boundaries that previously matched the Latin template.

 The measured Microsoft badge variant (2026-08-27 registration) is the first
 `tr`-corner mark and the first `long`-side scale basis: the pill tracks the
@@ -1758,7 +1822,8 @@ undersized the template by the aspect ratio (portrait carriers fell to
 its synthetic internal shapes knocked out - the holes are what separate it from
 any other bright rounded corner element (a plain white pill scores below the gate
 in the tests). It does not claim coverage of Microsoft's other documented icon,
-wording, or position variants.
+wording, or position variants beyond the separately gated beige `Made with AI`
+text pill.

 The 2026-08-27 rerun used the registered engine through
 `scripts/registered_mark_calibrate.py`, rather than a copied detector
@@ -1795,8 +1860,9 @@ box grows), not the whole `footprint_mask`. Baidu extends right to the corner ta
 and LiblibAI extends left to the triangle logo; both inherit every guard around
 that arithmetic.

-Doubao, Kling, and Samsung are deliberate sparse-mask exceptions because their alpha assets
-supply more information than a rectangle. Doubao's continuous top-hat response
+Doubao, Kling, and Samsung retain deliberate sparse masks for their original
+layouts because those alpha assets supply more information than a rectangle.
+Doubao's continuous top-hat response
 locates the mark, then `DoubaoEngine.footprint_mask` resizes the alpha to that same winning
 `match_box` and masks only the glyphs. The canonical 2048-pixel fixture has bright
 branch texture behind the bottom-edge wordmark; bounding the thresholded response,
@@ -1806,6 +1872,13 @@ with large triangular wedges. The aligned sparse mask keeps both frame edges
 untouched and still clears the detector. `force` cannot align a missing detection,
 so it retains the shared geometry-box fallback.

+The added outlined Doubao, stacked Yuanbao, Dola, WorkBuddy, Korean Samsung, and
+beige Microsoft layouts use solid detector-aligned footprints where a sparse
+synthetic glyph mask left visible outlines, sparkles, capsules, or shadows in
+rendered removal checks. Each layout keeps its own detection threshold and mask
+geometry; selecting one does not replace the calibrated behavior of the older
+layout under the same registry key.
+
 Doubao keeps the original low-saturation scan for pale corners. When enough of
 its corner is strongly colored and that scan falls below the strict gate, a
 wider color filter searches for white glyphs tinted by the background. The
@@ -1933,14 +2006,12 @@ exists.

 #### OpenArt wordmark

-Registered 2026-09-21 for a confirmed production miss reported through raiw-app:
-a free `visible`-mode run (2026-09-11, 1528x2712 portrait, Spaces uid
-`aefb1e07306f4a7280615a8daea01a39`) left a large, prominent OpenArt wordmark
+Registered 2026-09-21 after a `visible`-mode run left a large, prominent OpenArt wordmark
 (a bowtie/infinity icon followed by "OpenArt" in bold sans-serif, semi-
 transparent white) completely unremoved -- pixel-identical before and after.
 `identify` found nothing at all for that case: no C2PA, EXIF AI tag, or China
 AIGC/TC260 label, so the visible mark is the only known attribution route for
-a bare OpenArt export. The user rated the result 0 ("Didn't work").
+a bare OpenArt export.

 **This is the only registered mark placed at the FRAME CENTER rather than a
 corner.** Every other engine in this family anchors to one of the four corners
@@ -2004,10 +2075,14 @@ geometry, and until one exists the bare-arm question is settled by this
 measurement, not by gate tuning. The local calibration harness is not tracked.
 (cohorts A/B/C over the no-signal pool).

-The weak Jimeng pill detector lives in
-[`pill_engine.py`](../src/remove_ai_watermarks/pill_engine.py). It uses a
-synthetic silhouette for detection and a fixed top-left footprint. Its public,
-published real-world regression carrier is
+The Jimeng pill detector lives in
+[`pill_engine.py`](../src/remove_ai_watermarks/pill_engine.py). It combines the
+existing weak synthetic-outline path with a high-confidence contrast template
+for the measured solid `AI生成` label. A Jimeng provenance match can corroborate
+that verified label without the flat-background guard; the weak outline keeps
+both the provenance and background gates described above. Removal follows the
+matched label when verified and otherwise uses the fixed top-left footprint. Its
+public, published real-world regression carrier is
 `data/fixtures/visible/jimeng_pill/provider-published-example.jpg`; the direct
 detector scores the pill at 0.28 and the same image's Jimeng wordmark at 0.61.

@@ -2133,6 +2208,13 @@ image sizing, postprocessing, and the public engine
 interface. It delegates model execution to
 [`_internal/watermark_remover.py`](../src/remove_ai_watermarks/_internal/watermark_remover.py).

+The engine preserves source transparency for alpha-capable outputs. It captures
+alpha after EXIF orientation is applied and before resizing, then rejoins the
+unchanged alpha plane after restoring the RGB result to the original dimensions
+and applying postprocessing. This also handles palette and colour-key PNG
+transparency; diffusion itself still receives RGB. Regression coverage lives in
+`test_invisible_engine.py::TestTransparency`.
+
 `get_device` in that module answers only `cuda` or `cpu`. An `mps` or `xpu` answer
 would travel one frame to the same CUDA-only refusal while costing a device probe,
 and reporting it implied an Apple-silicon or Intel-GPU path that does not exist.
@@ -2162,27 +2244,27 @@ Regression coverage:

 ### CPU offload

-CPU offload is enabled only when requested. Nothing calls Diffusers'
-`enable_model_cpu_offload` any more -- that belonged to the deleted single-stage
-profiles. `--cpu-offload` sets both residency fields, and they do not travel the
-same distance: the face field is read by the shared base and so applies to every
-profile, while the global field is read by `QwenZImagePipeline` alone.
-`ChromaZImagePipeline` and `SdxlZImagePipeline` load their global stack with a
-plain `.to(device)` and never consult it.
+Nothing calls Diffusers' `enable_model_cpu_offload` any more -- that belonged to
+the deleted single-stage profiles. `--cpu-offload` sets both residency fields, and
+they do not travel the same distance: the face field is read by the shared base and
+so applies to every profile, while the global field is read by `QwenZImagePipeline`
+and `SdxlZImagePipeline`. `ChromaZImagePipeline` loads its global stack with a
+plain `.to(device)` and never consults it.

 `GLOBAL_OFFLOAD_PROFILES` in `watermark_profiles.py` is the list, and
 `global_offload_supported` the question; `WatermarkRemover._warn_if_global_offload_unsupported`
 warns ahead of the model load rather than at construction, because `auto` picks
 its engine per-image. An explicit `auto` preload uses the qwen-zimage fallback and
 therefore keeps global offload active before that resolution. Below
-`RESIDENT_FACE_MODEL_MIN_VRAM_GIB` the flag is a no-op for the other two profiles --
-the face stack was offloading anyway and the global one does not ask -- which is
-what the warning exists to say before a download starts.
+`RESIDENT_FACE_MODEL_MIN_VRAM_GIB` the flag is a no-op for `chroma-zimage` -- the
+face stack was offloading anyway and the global one does not ask -- which is what
+the warning exists to say before a download starts.

 Residency is otherwise chosen from the card's total VRAM, once per stack:
 `resolve_global_model_residency` gates the mandatory Qwen stack at
-`RESIDENT_GLOBAL_MODEL_MIN_VRAM_GIB` and `resolve_face_model_residency` gates the
-optional Z-Image stack at `RESIDENT_FACE_MODEL_MIN_VRAM_GIB`.
+`RESIDENT_GLOBAL_MODEL_MIN_VRAM_GIB`, `resolve_sdxl_global_residency` gates the
+SDXL stack at `RESIDENT_SDXL_GLOBAL_MIN_VRAM_GIB`, and `resolve_face_model_residency`
+gates the optional Z-Image stack at `RESIDENT_FACE_MODEL_MIN_VRAM_GIB`.

 Below the global floor, `_qwen_vram_config` streams the stack from disk, which is
 what makes a 20B model runnable on a consumer card. At or above it, streaming is
@@ -2199,6 +2281,43 @@ The resident config deliberately passes no `"disk"` value anywhere. DiffSynth la
 `disk_offload` once, from `offload_dtype`, so leaving the sentinel in place while
 pointing every device at CUDA would keep the meta-drop and re-read.

+Sequential SDXL offload receives the current CUDA device index explicitly, so
+its hooks use the same GPU as the residency probes, input tensors, and seeded
+generators. Omitting the index makes Diffusers select GPU 0 even after
+`torch.cuda.set_device(1)`. `test_streaming_uses_the_current_cuda_device` in
+`test_cpu_offload.py` covers automatic and forced streaming on nonzero devices.
+
+The SDXL stack (fp16 UNet, Canny ControlNet, both text encoders, the fp16-fix VAE)
+is ~8.8 GiB of weights, and the UNet and the ControlNet run together on every
+denoising step. Below the 12 GiB floor `_load_global` fuses the Lightning LoRA on
+the CPU and then calls `enable_sequential_cpu_offload`; the order matters, because
+the offload hooks stream their own copy of each weight and a later fuse would never
+reach it. Measured on an RTX 5060 Laptop (8 GiB, WSL2, 16 GiB system RAM) in
+October 2026, one 1672x941 global pass at strength 0.15:
+
+| Placement | Warm pass | Peak VRAM | Peak system RAM |
+|---|---|---|---|
+| resident `.to(device)` (previous behaviour) | did not finish in 10 min | 7.9 GiB used, spilling | -- |
+| `enable_model_cpu_offload` | 237 s | 9.2 GiB reserved on an 8 GiB card | 12.4 GiB |
+| group offload, block level, CUDA stream | 92 s, or hung | 12.4 GiB reserved | 15.6 GiB |
+| group offload, leaf level, CUDA stream | 7.8 s, or hung | 5.6 GiB | 15.7 GiB |
+| group offload, leaf level, `low_cpu_mem_usage` | 21.4 s | 5.6 GiB | 9.4 GiB |
+| `enable_sequential_cpu_offload` | 12.8 s | 3-4 GiB | 8.5 GiB |
+
+The resident and model-offload rows did not fail: under the Windows display driver
+CUDA falls back to shared system memory, so PyTorch reserved more than the card
+holds and the run crawled instead of raising. The streamed group-offload variants
+pin every weight in host memory up front; on a 16 GiB machine that left the run
+finishing or hanging depending on what else held RAM, which rules them out as a
+default. All completed placements produced bit-identical images.
+
+Fusing on the CPU instead of on CUDA rounds 788 of the UNet's 3256 fused tensors
+differently, by at most one fp16 ulp (1.2e-4). On the pass above that moved the
+output by a mean of 0.16 levels (PSNR 54.3 dB against a CUDA-fused, equally
+streamed run). Cards at or above the floor keep the CUDA fuse and therefore the
+previous pixels; only a streamed run, automatic or forced with `--cpu-offload`,
+sees the difference.
+
 Regression coverage:

 - [`test_cpu_offload.py`](../tests/test_cpu_offload.py)
@@ -14,6 +14,7 @@ through the installed CLI or public Python API.
 | `openai-api` | OpenAI SynthID | Official Content Provenance API | Images |
 | `microsoft-api` | Microsoft InvisMark | Azure Content Provenance Detection API | Images |
 | `gemini-web` | Google SynthID | Signed-in Gemini web account | Images and videos |
+| `synthid-web` | Partner SynthID | SynthID Detector at synthid.com, real Chrome | None recorded |
 | `meta-web` | Meta Content Seal | Anonymous Meta identification page | Images |
 | `microsoft-web` | Microsoft InvisMark | Public Azure AI validation page | Images |
 | `openai-web` | OpenAI SynthID | Public web verifier | Images |
@@ -23,26 +24,46 @@ through the installed CLI or public Python API.
 The default selection order is API first, then Web. OpenAI therefore plans
 `openai-api` before `openai-web`, and Microsoft plans `microsoft-api` before
 `microsoft-web`. Google and Meta currently have no usable API adapter in this
-tooling, so their Web surfaces are primary rather than fallback routes.
+tooling, so each of their Web surfaces is primary. Google has two: `gemini-web`
+and `synthid-web`.

 Inspect the machine-readable order before preparing a batch:

 ```bash
 uv run python scripts/provider_oracles.py plan openai --json
 uv run python scripts/provider_oracles.py plan microsoft
+uv run python scripts/provider_oracles.py plan google
 ```

-API-first is a selection policy, not automatic failover. Use Web only when the
-provider, requested media, or local configuration has no usable API route, or
-when the operator explicitly requests a second surface. An API refusal, quota
-response, transport failure, or indeterminate verdict is recorded as-is and
-never silently triggers a Web upload.
+A request to use a provider covers every configured surface of that provider,
+including Web, without a second confirmation. API-first is still the order to
+try, not a lock on the other surfaces. An API refusal, quota response, transport
+failure, or indeterminate verdict is recorded as-is. It does not silently
+upload the same file to Web; run that surface as its own acknowledged check and
+keep the two verdicts separate.

-The old Google SynthID Detector portal does not work. On 2026-09-07, the
-official `labs.google/synthid` entry point redirected a signed-in browser to a
-Google 404; on 2026-09-23 it redirected to
-`deepmind.google.com/frontiers/synthid`, which also returned 404. Historical corpus rows may still name `synthid-portal`; that value
-describes how an old result was obtained and is not an operational surface.
+The old Google SynthID Detector portal at `labs.google/synthid` does not work.
+On 2026-09-07 that entry point redirected a signed-in browser to a Google 404;
+on 2026-09-23 it redirected to `deepmind.google.com/frontiers/synthid`, which
+also returned 404. Historical corpus rows may still name `synthid-portal`; that
+value describes how an old result was obtained and is not an operational surface.
+
+The public SynthID Detector is a different site, [synthid.com](https://synthid.com/),
+and it is the `synthid-web` surface. On 2026-10-17 the page title was
+`SynthID Detector`. After the terms dialog, the uploader was present without a
+completed sign-in in that session. The Account menu still offered Sign in, and
+no file was submitted, so this check does not show whether a result requires an
+account. The live FAQ named the partners as Google DeepMind, Nvidia, OpenAI,
+and Kakao, and it did not list Apple. It defined two settled labels:
+`Watermark detected` and `Watermark not detected`. The file input accepted
+`.jpg`, `.jpeg`, `.png`, `.bmp`, `.webp`, `.avif`, `.heic`, `.heif`, `.tiff`,
+`.tif`, `.gif`, `.wav`, `.mp3`, `.ogg`, `.flac`, `.aac`, `.m4a`, `.mp4`,
+`.mov`, and `.webm`. This tooling still prepares only the project's existing
+JPEG, PNG, WebP, MP4, MP3, and WAV suffixes. The portal's additional terms
+forbid using the page as an automated API, bulk verification, and quota evasion,
+so `run-web` does not drive it. Use the real Chrome workflow below. A
+`Watermark detected` result is partner SynthID, not a Gemini-only or OpenAI-only
+verdict.
 Google's newer Search verification entry point did not provide an independent
 oracle allowance in the session tested on 2026-09-13. Lens accepted the
 metadata-stripped, previously Gemini-positive image
@@ -161,9 +182,12 @@ Adobe Inspect and local C2PA, DWT-DCT, TrustMark, and research detectors are
 useful evidence readers, not interchangeable provider pixel-watermark oracles.

 The surface capability and this project's evidence coverage are separate. The
-current Google, Microsoft, and Meta pages accept image, video, and audio inputs;
-the OpenAI API accepts image and audio. A media type does not become certified
-merely because a surface accepts it. Run `list --json` to read both fields.
+current Google, Microsoft, and Meta pages, including `https://synthid.com/`,
+accept image, video, and audio inputs; the OpenAI API accepts image and audio.
+A media type does not become certified merely because a surface accepts it.
+`synthid-web` therefore lists image, video, and audio as accepted media and
+records no project evidence until a batch is verified. Run `list --json` to read
+both fields.

 ## Execution slots

@@ -193,6 +217,13 @@ default; `--slots PATH` explicitly overrides it.
      "google_account_index": 2,
      "network_label": "office"
    },
+    {
+      "name": "synthid-personal",
+      "surface": "synthid-web",
+      "account_label": "google-account-a",
+      "browser_profile": "Your Chrome",
+      "network_label": "home"
+    },
    {
      "name": "openai-web-thordata-us",
      "surface": "openai-web",
@@ -261,7 +292,9 @@ default; `--slots PATH` explicitly overrides it.
 ```

 A routed batch selects one slot explicitly. A Gemini slot must name the Google
-account, its real Chrome profile, and its nonnegative `/u/N/` account index; an
+account, its real Chrome profile, and its nonnegative `/u/N/` account index. A
+SynthID Detector slot names the account and the same real Chrome profile, and it
+does not use `google_account_index`. An
 API slot must name its account or project and cannot name a browser profile. The
 tool records the network route that the operator has already established.
 Non-Google Web slots may name `proxy_url_env`; Playwright reads the secret proxy
@@ -279,9 +312,10 @@ again for files or accounts already covered by the task.
 Submit each frozen artifact at most once per slot in that pass. Keep every
 verbatim refusal, inconclusive reply, and disagreement between accounts. Stop
 when the planned checks finish or no authorized usable slot remains, and report
-unsubmitted rows. A new login, another provider, a new network route, or a new
-pass after the bounded plan is exhausted needs authorization unless already
-requested. Candidate pixels remain frozen throughout account continuation.
+unsubmitted rows. A new login, another provider, or a new network route needs
+authorization. The oracles and slots already configured for the requested
+provider may be used again, including their Web surfaces, without a second
+confirmation. Candidate pixels remain frozen throughout account continuation.

 Each `openai-api` slot must name exactly one `api_key_env`. Put the corresponding
 values in the gitignored local `.env`, never in the slot JSON:
@@ -385,18 +419,23 @@ uv sync --extra dev
 uv run playwright install chromium
 ```

-Google is intentionally excluded from the standalone runner because its oracle
-requires the user's existing authenticated Chrome session. Invoke the project
-`provider-oracles` skill to prepare the batch, select the named real Chrome
-profile and `/u/N/` Google account, upload each prepared artifact, capture the
-settled SynthID tool outcome, and record it.
+Google is intentionally excluded from the standalone runner. Both `gemini-web`
+and `synthid-web` use the user's existing real Chrome profile, and the SynthID
+Detector terms forbid driving that page as an automated API. Invoke the project
+`provider-oracles` skill to prepare the batch and record the settled outcome.
+For Gemini, select the named profile and `/u/N/` account. For SynthID Detector,
+open `https://synthid.com/` in that profile. The portal has no `/u/N/` index.
+Accept the terms dialog when it is shown, attach one prepared upload, and read
+the explicit `Watermark detected` or `Watermark not detected` label. If the page
+requires sign-in and that profile has no session, record the row unsubmitted
+and do not start a new login.

-Ask the direct question stored in the batch manifest: `Does this
+For Gemini, ask the direct question stored in the batch manifest: `Does this
 image/video/audio contain a SynthID watermark? Use the SynthID verification
 tool only; do not infer from visible content, logos, text, or metadata.` A
 general question about whether Google AI created or edited the media is not
 equivalent: it may elicit origin or visual-content reasoning without a SynthID
-verdict.
+verdict. SynthID Detector has no prompt; its settled labels are the verdict.

 OpenAI Web, Microsoft Web, and Meta run entirely through a fresh isolated
 Playwright browser. A direct-route example:
@@ -24,9 +24,10 @@ The package version appears in:

 - `pyproject.toml`;
 - `src/remove_ai_watermarks/__init__.py`;
+- both `version` fields in `server.json`;
 - the root package entry generated in `uv.lock`.

-Update the first two, then refresh the lock file with uv. Do not edit a
+Update the first three, then refresh the lock file with uv. Do not edit a
 line-number-specific location in `uv.lock`; its package order changes.

 Before choosing a version, verify it is absent from git tags, GitHub Releases,
@@ -159,7 +160,10 @@ The Agent Skill and Claude plugin live in this repository:
  is the Claude plugin manifest, with source `./skills` so an install copies
  only the skill tree;
 - [`.claude-plugin/marketplace.json`](../.claude-plugin/marketplace.json)
-  makes the repository a Claude Code marketplace.
+  makes the repository a Claude Code marketplace;
+- [`.codex-plugin/plugin.json`](../.codex-plugin/plugin.json) is the Codex
+  plugin manifest, pointing at the same `./skills/` tree, with its icon in
+  [`assets/icon.svg`](../assets/icon.svg).

 They are not a fifth required release surface. A CLI change that affects
 command routing, extras, exit codes, mark keys, or the intended-use boundary
@@ -215,6 +219,26 @@ the catalog has not accepted the initial submission. The catalog-specific
 submission state and links live in
 [`docs/agent-skill.md`](agent-skill.md#where-the-skill-can-actually-be-listed).

+## MCP registry
+
+[`server.json`](../server.json) lists the PyPI package in the
+[official MCP registry](https://registry.modelcontextprotocol.io/) as
+`io.github.wiltodelta/remove-ai-watermarks`, started as
+`uvx remove-ai-watermarks mcp`. The registry accepts it only after PyPI serves a
+release whose description, the README, carries the `mcp-name:` marker, and
+`tests/test_mcp_server.py` pins `server.json` to the package version. The listing
+is not a release surface that blocks a release; after PyPI publishes a version:
+
+1. `mcp-publisher login github` with an account that owns `wiltodelta`;
+2. `mcp-publisher validate` and then `mcp-publisher publish` from the
+   repository root;
+3. read the entry back from
+   `https://registry.modelcontextprotocol.io/v0/servers?search=io.github.wiltodelta/remove-ai-watermarks`
+   and check its version.
+
+[`glama.json`](../glama.json) names the maintainer for Glama's claim flow.
+Client setup and the tool contract are in [the MCP server guide](mcp-server.md).
+
 ## Release verification

 Forensic transports are versioned independently from the package. Before publishing
@@ -33,22 +33,28 @@ This page is an index. Each finding lives in its canonical page.
 | Dependencies | TrustMark extra on every supported Python (`trustmark>=0.9.2`) | [installation.md](installation.md) |
 | Live captures | Gemini Omni, Grok, Vidu, Wan, Apple, Google Photos, Nano Banana 2 Lite and Gemini Omni Flash through the API; Runway, Higgsfield and Dreamina | [watermarking-landscape.md](watermarking-landscape.md#live-captures-2026-09-23-and-2026-09-24) |

-## Open items
+## Follow-up status

- **Google Photos Reimagine manifests.** Magic Editor, and so Reimagine, is not
-  offered in Google Photos on iOS, so no Reimagine file is captured. The
+The original follow-up list now mixes completed experiments, deliberately
+excluded claims, and acquisition gaps. Each row below names its current status
+as of 2026-10-01 so this section can serve as the research backlog without
+discarding the measured results.
+
+- **Closed - Google Photos Reimagine manifests.** Magic Editor, and so
+  Reimagine, is not offered in Google Photos on iOS, so no Reimagine file is
+  captured. The
  question it was meant to answer is settled without one: Google's checker found
  SynthID in all four Photos AI edits tested on 2026-09-25, none of which
  records SynthID in its manifest.
- **Vidu and Wan marks on more samples.** Both are registered from one real
-  export each (`vidu` video, `wan` image); their gates are provisional until a
+- **Open - Vidu and Wan marks on more samples.** Both are registered from one
+  real export each (`vidu` video, `wan` image); their gates are provisional until a
  wider cohort is measured, and a Wan video label is not yet captured. Public
  Vidu feed videos (2026-09-26) carry the TC260 label naming ShengShu, and one
  names Tongyi Yunqi as producer with Vidu as propagator, but none shows the
  visible `Vidu` mark; Wan's public demo videos show no mark either. Both marks
  appear only on account downloads.
- **TC260 practice guides.** The audio placements from TC260-PG-202510A are
-  now read (the guide PDFs load in a browser from
+- **Partially complete - TC260 practice guides.** The audio placements from
+  TC260-PG-202510A are now read (the guide PDFs load in a browser from
  `tc260.org.cn/tc260/sjzn/list.shtml`, not to scripted requests). The text-file
  guide (TC260-PG-20258A) covers documents this image and video tool does not
  process. The security-protection guide (TC260-PG-202511A) records a
@@ -66,15 +72,25 @@ This page is an index. Each finding lives in its canonical page.
  point; changing one `ProduceID` digit fails it. MiniMax's own spellings depart
  from the guide: `TBSData.Type` is `LabelMataData` (the guide's `Md`) and
  `Binding` with `BType` `0` (its `Bnd`); the second signature verifies over the
-  compact JSON of the `Bindings` entry, also on all four. That entry's SM3
-  content hash matched neither the whole file nor the file without its label,
-  and it names no `CntSel`, so the content binding is unverified. Other producers fill the same
+  compact JSON of the `Bindings` entry, also on all four. TC260-PG-202511A
+  appendix A.2 was re-read from the
+  [official PDF](https://www.tc260.org.cn/tc260/sjzn/202508/63ab55a3265f45c7865e9b8be15d4a29/files/dc379ec776f94abc8f1806affa1cda35.pdf)
+  on 2026-10-01. MiniMax omits its `CntSel`, and none of the guide-derived
+  byte selections reproduced its SM3 content hash on four measured videos, so
+  the content binding remains unverified
+  ([module internals](module-internals.md)). Other producers fill the same
  fields with opaque vendor strings (Qwen a base64 digest, Vidu 32 hex digits,
  Wan and HappyHorse `L-`/`K-` prefixed values), not `SecurityData`. The key sits
  in the file it signs, so a check proves authenticity only against a pinned
  MiniMax key. `identify` now verifies it with a built-in SM3/SM2 and a pinned
-  MiniMax key ([module internals](module-internals.md)).
- **Experiments.** Chroma restoration after regeneration (arXiv:2508.21072) was
+  MiniMax key. The remaining work is
+  MiniMax content-binding verification if its byte-selection rule becomes known,
+  and Honor's unsupported RSA/content form if an original export and signing
+  contract become available.
+- **Partially complete - experiments.** Five of the six follow-ups are closed:
+  Z-Image-Turbo, Perth and PixelSeal, re-watermarking, VACE/ROSE, and newer
+  frozen-backbone probes. Chroma
+  restoration after regeneration (arXiv:2508.21072) was
  measured on 2026-09-27 and is not shipped: transplanted source chroma drew one
  explicit "inconclusive" and two unclear Gemini answers where the untouched
  outputs were all clean, and ghosts redrawn objects
@@ -126,12 +142,20 @@ This page is an index. Each finding lives in its canonical page.
  profile; retain its harness and local evidence for future comparison only.
  The existing source-conditioned Z-Image face-repair stage remains part of
  `qwen-zimage`, `sdxl-zimage`, and `chroma-zimage`.
-  Still open: linear probes on newer frozen backbones for `classify` and RAVEN
-  view-synthesis removal
-  (arXiv:2601.08832, no code released as of 2026-09-24).
+  The pre-registered modern-backbone campaign found that neither frozen
+  SigLIP2 Base nor PE-Core B could match the production model's combined photo,
+  AI, and FLUX constraints, so the replacement hypothesis is closed
+  ([classifier research](ai-generated-image-classifiers.md#modern-frozen-backbone-campaign-closed-2026-09-27)).
+  Only RAVEN view-synthesis removal remains blocked. RAVEN's
+  [official repository](https://github.com/fahadshamshad/raven-watermark-removal)
+  now exists, but its 2026-08-08 README is still a release placeholder: the
+  authors say they will publish code after their SynthID testing and Google
+  Security Bounty submission. There is no implementation to evaluate as of
+  2026-10-17, when the repository still held only that README (arXiv:2601.08832).
  MiniMax-Remover (CC-BY-NC-4.0 weights) and ProPainter (S-Lab License,
-  non-commercial) are excluded on license.
- **New C2PA signers without a vendor row.** The
+  non-commercial) are deliberately excluded on license and are not backlog
+  items.
+- **Open - new C2PA signers without a vendor row.** The
  [C2PA conformance list](https://github.com/c2pa-org/conformance-public/blob/main/conforming-products/conforming-products-list.json)
  names generator certificates `Amazon Bedrock`, `Getty Images
  AI-Generated Image` and `AI-Modified Image` (and the iStock pair), `Jasper`,
@@ -148,7 +172,7 @@ This page is an index. Each finding lives in its canonical page.
  carrying C2PA: their marketing and model-page images are unsigned. A vivo
  X300 camera capture carries a vivo device manifest (`digitalCapture`) that
  the bundled reader cannot open (`unknown algorithm`).
- **Chinese phone galleries.** Huawei's help pages say the Xiaoyi photo-edit AI
+- **Open - Chinese phone galleries.** Huawei's help pages say the Xiaoyi photo-edit AI
  watermark can be switched off; vivo's Album setting is already covered.
  The public-web follow-up below found Honor YOYO metadata and additional
  primary-source labeling evidence, but not a controlled gallery export set. A live
@@ -159,11 +183,12 @@ This page is an index. Each finding lives in its canonical page.
  of an unedited control, an ordinary crop, and a generative edit are still needed
  per app before selecting an AI attribution rule. Record the device, OS/app
  version, region, operation, and original download path with each sample.
- **Claims not supported by a primary source.** Kling embedding SynthID
+- **Closed - claims not supported by a primary source.** Kling embedding SynthID
  (repeated by remover blogs; Kuaishou is not in Google's adopter list),
  Hailuo pixel watermarks, and platforms stripping C2PA on upload. None is used
  by the library.
- **Google Photos small edits (further work deferred, 2026-10-01).** The
+- **Partially complete - Google Photos small edits (further work deferred,
+  2026-10-01).** The
  maintainer deferred further photographic tiny Magic Eraser captures and
  verification retries for the inconclusive small edits. Resume only on an
  explicit request; retain the existing evidence and uncertainty. On
@@ -193,8 +218,13 @@ This page is an index. Each finding lives in its canonical page.
  negative and one inconclusive reply. The quota refusal, synthetic capture,
  pixel comparison, mode caveat, and hash-bound outcomes are recorded in
  [synthid.md](synthid.md).
- **Captures blocked by region or device.** Doubao, Jimeng, Samsung. Higgsfield,
-  Runway and Dreamina were captured on paid plans on 2026-09-24.
+- **Partially complete - direct captures blocked by region or device.** Direct,
+  controlled captures remain blocked for Doubao, Jimeng, and Samsung. The
+  public-web follow-up below later found one real marked Doubao video and
+  validated its visible mark and TC260 metadata, but the file remains local
+  because publication has not been cleared; it is not a controlled account
+  capture or a redistributable fixture. Higgsfield, Runway, and Dreamina were
+  captured on paid plans on 2026-09-24.

 ### Public-web sample search, 2026-10-01 UTC

@@ -251,10 +281,16 @@ provider exports. No third-party photograph was added to tracked fixtures.
  [Getty's generator page](https://www.gettyimages.com/ai/generation),
  [iStock's generator page](https://www.istockphoto.com/ai/generation), and
  [Reface's product site](https://reface.ai/) did not carry locally recognized
-  C2PA. Jasper explicitly distinguishes input/mask/output examples, but an
-  unsigned web derivative cannot establish its signed export's manifest shape.
+  C2PA. A 2026-10-01 recheck of Jasper's current official Image landing page
+  reached its pre-rendered endpoint outputs directly and again found no C2PA.
+  Jasper explicitly distinguishes input/mask/output examples, but an unsigned
+  web derivative cannot establish its signed export's manifest shape.
  Conformance certificates alone still do not justify fabricating that shape.
- **Visible-image/video gaps remain.** The downloaded RunningHub workflow
+- **Visible-image/video provider-original gaps remain.** Every registered mark
+  already has a publication-safe synthetic detector fixture in the
+  [visible-mark gallery](../data/fixtures/visible/README.md); the sample ledgers
+  now point to those fixtures and distinguish them from provider evidence. The
+  downloaded RunningHub workflow
  illustrations are explanatory/branding assets, not verified output examples;
  Liblib API documentation's cat illustration has no visible historical wordmark.
  A public [Inkpainting model page](https://www.liblib.art/modelinfo/2acd393c23924ff9b5363c9c1ec9129f)
@@ -11,19 +11,21 @@ The `visible` command registers these mark keys:
 | Key | Mark | Expected area | Important limit |
 | --- | --- | --- | --- |
 | `gemini` | Google Gemini visible watermark (sparkle) | Usually bottom right | Detection includes a false positive gate. |
-| `doubao` | `豆包AI生成` | Bottom right | Vendor specific text detector. |
+| `doubao` | `豆包AI生成` | Bottom right | Covers filled and outlined text layouts; the weak outlined layout requires matching Doubao provenance. |
 | `jimeng` | `★ 即梦AI` | Bottom right | Vendor specific text detector. |
 | `qwen` | `千问AI生成` or the three-lobe Qwen symbol | Bottom right | Separate strict templates cover the text and Qwen Create symbol variants. |
 | `wan` | Wan logo and `Wan` wordmark | Bottom right, hugging the corner | Provisional: one real Wan 2.7 Pro export. Requires NCC >= 0.55 and a match that ends within 0.015 of the short side from both edges. |
-| `kling` | `可灵AI 3.0` or `KlingAI 3.0` | Bottom right | Separate calibrated silhouettes cover the older CJK and current IMAGE 3.0 Latin variants. |
-| `yuanbao` | `元宝` over `AI生成` | Bottom right | Standard two-line variant only. |
-| `samsung` | `✦ Contenuti generati dall'AI` | Bottom left | Calibrated for the Italian text variant. |
+| `kling` | `可灵AI 3.0` or `KlingAI 3.0` | Bottom right | Separate calibrated silhouettes cover the older CJK and current IMAGE 3.0 Latin variants; a glyph-detail gate rejects plain UI boundaries. |
+| `yuanbao` | `元宝` over `AI生成` | Bottom right | Separate strict and provenance-corroborated templates cover the measured two-line rasterizations. |
+| `dola` | `Dola AI` | Bottom right | Text detector and detector-aligned wordmark footprint. |
+| `workbuddy` | `AI生成` over `WORKBUDDY>_` | Bottom right | Text detector and detector-aligned stacked-label footprint. |
+| `samsung` | Italian or Korean Galaxy AI disclosure | Bottom left | Independently gated text layouts cover `Contenuti generati dall'AI` and `AI로 생성한 콘텐츠`. |
 | `runninghub` | `RunningHub AI生成` | Top left | Strict visual and position gates. |
 | `baidu` | `百度 AI生成` | Bottom right | Detector and extended removal footprint. |
 | `liblib` | `LiblibAI` | Bottom center | Historical wordmark variant; includes a minimum image size gate. |
 | `liblib_pill` | Compact `AI生成` pill | Top left | Auto removal requires LiblibAI metadata or the bottom-center wordmark, plus a flat background under the pill. |
-| `microsoft` | One Microsoft white AI-badge variant | Top right | Strict uses the visual gate; auto can use Microsoft provenance for the measured [relaxed gate](module-internals.md#visible-mark-removal). Other documented icon, text, and position variants are not covered. |
-| `jimeng_pill` | `AI生成` pill | Top left | Weak detector with additional product and background gates. |
+| `microsoft` | Microsoft white AI badge or beige `Made with AI` pill | Top right | Each layout has its own strict visual gate; auto can use Microsoft provenance for the original badge's measured [relaxed gate](module-internals.md#visible-mark-removal). |
+| `jimeng_pill` | `AI生成` pill | Top left | A verified label layout can bypass the flat-background guard only with Jimeng provenance; the weak outline detector retains the product and background gates. |
 | `generic_ai_label` | Bare `AI生成` text, no vendor wordmark | Bottom right | Brand-less TC260 compliance-stamp fallback (vivo and Xiaomi Gallery AI-edit output confirmed; see [module-internals.md](module-internals.md#generic-bare-ai生成-label)). Font-rendered synthetic asset; **calibrated on a synthetic corpus only**, not a real captured generator screenshot -- treat as weaker evidence than the vendor-tuned detectors above. |
 | `openart` | `OpenArt` wordmark (bowtie/infinity icon + brand name) | Frame center, not a corner | Confirmed from one customer-reported production case; no known metadata signal, so the visible mark is the only attribution route. **UNCALIBRATED**: procedurally reconstructed asset, no captured corpus -- see [module-internals.md](module-internals.md#openart-wordmark). |

@@ -93,6 +95,8 @@ The inspection and stripping code handles signals in these groups:
 - exact app-export provenance and AIGC disclosures from supported
  ByteDance-family products, with product-only provenance excluded from the
  generated-image verdict;
+- Kling's H.264 `user_data_unregistered` SEI tag (`kling-ai`) in MP4/MOV
+  video, blanked in place by `video metadata --remove` without re-encoding;
 - IPTC AI disclosure fields;
 - PNG text chunks and embedded generation parameters, including keyed
  `workflow` and `prompt` tags in MP4/MOV metadata lists;
@@ -298,7 +302,7 @@ not a universal clean verdict.
 | Luma AI | None registered | No registered pixel decoder | PNG text generator tags (Uni-1) |
 | ByteDance generators | Doubao and Jimeng marks | No registered pixel decoder | TC260 AIGC, supported C2PA, and exact app-export AIGC disclosures |
 | Qwen | Qwen mark | No registered pixel decoder | TC260 AIGC |
-| Kling AI | Kling AI image and video marks | No registered pixel decoder | TC260 AIGC |
+| Kling AI | Kling AI image and video marks | No registered pixel decoder | TC260 AIGC and the video SEI tag |
 | Vidu (ShengShu) | Vidu video mark | No registered pixel decoder | TC260 AIGC in the MP4 `AIGC` key |
 | Runway | None measured on paid-plan output | No registered pixel decoder | C2PA signed `RUNWAY AI, INC.` on its own models; third-party models keep their vendor's manifest |
 | Amazon Bedrock (Nova Canvas, Nova Reel) | None | Invisible watermark readable only through the Bedrock API | C2PA signed `Amazon Web Services, Inc.`, claim generator `Amazon Bedrock` |
@@ -200,7 +200,7 @@ route in this campaign. Mechanism detail stays in
 | --- | --- | --- |
 | Gowal et al., [arXiv:2510.09263](https://arxiv.org/abs/2510.09263) | Post-hoc encoder `f` / decoder `g`. Detection logit is not payload recovery. SynthID-O (partner variant) 136 bits at 512x512. TPR at 0.1% FPR 99.98% aggregated, 99.97% on the hardest spatial crop+resize. Trains against sampled semantics-preserving transforms, including weak VAE regeneration. Production decoder unpublished | Matches the architecture we treat as keyed `x' = x + g(x)`. Explains why a two-pixel crop kills `pipeline_lattice` but not the official oracle, and why a 22.3 dB foreign VAE still reads `detected` |
 | DeepMind [US12094474B1](https://patents.google.com/patent/US12094474B1/en) and continuation [US20250149048A1](https://patents.google.com/patent/US20250149048A1/en) | Residual U-Net encoder, separate decoder, optional key, encoder/decoder ensembles that need not recognize each other | Constraint, not a recipe. Ensemble non-recognition is why one recovered Google tile cannot be reused as an OpenAI detector (S4) |
-| Dathathri et al., [Nature 634:818-823 (2024)](https://www.nature.com/articles/s41586-024-08025-4) | SynthID-Text: tournament sampling of LLM tokens, open-source | Different system. Image/audio/video remain proprietary |
+| Dathathri et al., [Nature 634:818-823 (2024)](https://www.nature.com/articles/s41586-024-08025-4) | SynthID-Text: tournament sampling of LLM tokens, open-source. Anthropic's [Claude text watermark](https://www.anthropic.com/news/claude-text-watermark) (announced 2026-08-14 for future Claude models, updated 2026-09-01) "is a version of the SynthID-Text approach", with its detection API in private preview | Different system. Image/audio/video remain proprietary |
 | OpenAI, [advancing content provenance](https://openai.com/index/advancing-content-provenance/) (2026-05-19, audio 2026-07-31) and [content provenance API](https://developers.openai.com/api/docs/guides/content-provenance) | ChatGPT / API / Codex images carry C2PA plus SynthID. Audio from 2026-07-31. `POST /v1/content_provenance_checks`. `not_detected` does not rule out another vendor | This is the oracle. C2PA and SynthID are independent entries. Do not abuse the endpoint as an adaptive reverse-engineering loop |

 ### Keyless presence detectors in the literature
@@ -229,12 +229,20 @@ regeneration is already in [synthid.md](synthid.md#23-removal-attacks-and-forens
 | Goonatilake and Ateniese, [arXiv:2605.09203](https://arxiv.org/abs/2605.09203) | Six removers all leave a forensic residue a ResNet-50 sees at >98% TPR @ 1% FPR | Applies to UnMarker, Zhao's WatermarkAttacker, CtrlRegen+ | Defeating a provider oracle is not deniability. This is the product remainder |
 | An et al., [arXiv:2401.08573](https://arxiv.org/abs/2401.08573) (WAVES, ICML 2024) | 26 attacks on StegaStamp, Stable Signature, Tree-Ring. Regeneration, not JPEG, is the attack that matters. StegaStamp TPR@1%FPR 1.00 to 0.01; Tree-Ring 0.99 to 0.12 | Open watermarks | Protocol. Our blur-sigma-7 and 16-32 scramble are closer to WAVES "distortion" than to regeneration |
 | Bulychev et al., [arXiv:2605.16796](https://arxiv.org/abs/2605.16796) (Watermarks Attack Watermarks) | Apply a second watermark, usually with the victim's own post-processing encoder and a different message. On MS-COCO, its end-to-end pipeline lowers Video Seal victim bit accuracy from 0.998 to 0.520 (TPR@1%FPR 0.046) and Pixel Seal from 0.999 to 0.514 (TPR 0.029); the same-method attacker message recovers at 0.99 bit accuracy for both schemes. Quality is reported as an eight-metric normalized composite, not a raw PSNR operating point | Eight open image schemes, not production SynthID; Video Seal is tested on images | Directly motivated the [local open-encoder replication](#re-watermarking-on-open-encoders-2026-09-27). Its MIT [code](https://github.com/MariaBulychev/Watermarks-Attack-Watermarks/tree/0444f52ad972da9e009555f2dc1028d9abe0610e) calls external encoder packages and weights; the top-level license does not license those dependencies |
+| Yang et al., [arXiv:2609.01249](https://arxiv.org/abs/2609.01249) (One Prompt Is Enough) | One reconstruction prompt to a closed image editor (GPT Image 1, 1.5 and 2; Nano Banana, Pro and 2) launders the payload over 1,800 outputs; OpenAI models disrupt it most | DwtDct, DwtDctSvd and RivaGAN only. "SynthID was not tested" | Regeneration through a provider's own model. It breaks the scheme our DWT-DCT detector reads, and says nothing about production SynthID, which the same editors add to their output |
+| Yao et al., [arXiv:2610.11290](https://arxiv.org/abs/2610.11290) (ProxyEraseAgent) | Ranks a pool of open decoders by calibrated response and attacks the strongest as proxy boundary estimators; 94.8% success over 11 systems against 79.5% for the best fixed Regen-VAE | Open schemes, TrustMark and VideoSeal among them. Not SynthID | Needs a decoder that responds to the mark. No public decoder responds to production SynthID, so it does not transfer to an oracle-only target |
+| Bao et al., [arXiv:2609.08213](https://arxiv.org/abs/2609.08213) (DRIFT) | Deflects the diffusion trajectory. The adaptive variant searches a strength ladder per image for "the first verifier-rejected rung" from binary verifier feedback; 98-100% success on nine marks | In-generation diffusion marks (Tree-Ring family, Gaussian Shading and others). Not SynthID | The ladder stop is our [attack ladders](#attack-ladders-openai-oracle) run per image. Against a provider verifier that is the adaptive query loop its terms forbid, so it applies to local oracles only |
+| Ye et al., [arXiv:2610.03166](https://arxiv.org/abs/2610.03166) (LiBRA) | White-box latent optimization with the key and decoder, steering decoding confidence toward chance from both sides so the mark is not left inverted but detectable; Stable Signature detection 1.00 to 0.03 | Released checkpoints of Stable Signature, the YU1 GAN fingerprint, HiDDeN and MBRS | Needs the decoder, so it is out of reach here. The inverted-mark point applies to any bit-flip attack |
+| Aistov et al., [arXiv:2610.02010](https://arxiv.org/abs/2610.02010) (Latent Frequency Masking, ICDM 2026) | Replaces selected Fourier coefficients of the latent; [code](https://github.com/alvitrioliks/lfm) | Stable Signature, Tree-Ring, RingID, Gaussian Shading (resists), Maxsive, METR. Not SynthID | In-generation latent marks. Not the post-hoc encoder family Gowal describes |
+| Li et al., [arXiv:2609.31693](https://arxiv.org/abs/2609.31693) (Disentangle and Drop) | Splits the image into a grayscale carrier and a residual branch, then suppresses the residual, with one shared operating setting | SSL, DwtDct, DwtDctSvd, RivaGAN, StegaStamp, Stable Signature, Tree-Ring | Open schemes only |
+| Yang et al., [arXiv:2609.31722](https://arxiv.org/abs/2609.31722) (Push-Pull) | Key-free and decoder-free at deployment; average BER 0.3958 at 31.07 dB PSNR on a four-method sweep | Seven open methods. Not SynthID | A BER near 0.40 leaves much of the payload intact, as with SPFM-Net below |
+| Ding et al., [arXiv:2609.34003](https://arxiv.org/abs/2609.34003) (DeMark, audio) | Query-free black box that suppresses the time-frequency artifacts the embedding leaves; average attack success 0.92 and 0.96 on two speech sets | AudioSeal, WavMark, Timbre, AudioMarkNet. Not audio SynthID | A candidate cohort for the pinned AudioSeal oracle in the [benchmark kernel](benchmark-kernel.md). No code located |
 | Wen et al., [arXiv:2305.20030](https://arxiv.org/abs/2305.20030) (Tree-Ring) | In-generation Fourier pattern in initial noise | Not SynthID | Kill is DDIM inversion plus latent wipe, not a pixel scramble. Lin and Juarez [arXiv:2506.10502](https://arxiv.org/abs/2506.10502) (USENIX 2025) remove it from public knowledge |
 | Fernandez et al., Stable Signature, ICCV 2023 | Fine-tuned VAE decoder | Not SynthID | Regeneration that replaces the decoder is in-family. Our foreign VAE at 22.3 dB did not kill OpenAI SynthID, which is the expected mismatch |
 | Google to [The Verge, 2026-04-14](https://www.theverge.com/ai-artificial-intelligence/911579/google-synthid-ai-watermarking-system-reverse-engineered) | reverse-SynthID does not systematically remove SynthID | Google spokesperson on the spectral bypass | Agrees with our V4 audit and with Bypass: reconstruction or decoder confusion, not a payload delete |

 Quiet pixel-only removal is not a published result on production
-OpenAI or Google SynthID (still true in the 2026-09-27 literature sweep).
+OpenAI or Google SynthID (still true in the 2026-10-17 literature sweep).
 Against open encoders the fidelity band has moved: WmForger
 ([arXiv:2510.20468](https://arxiv.org/abs/2510.20468)) and Hide&Seek
 ([arXiv:2603.01067](https://arxiv.org/abs/2603.01067)) report roughly 30 dB,
@@ -575,7 +575,11 @@ The SynthID decoder is proprietary and not released:
 > -- Gowal et al., arXiv:2510.09263

 There are no released payload-decoder weights or public algorithm. Google
-provides verification in Gemini. The old limited SynthID Detector portal does
+provides verification in Gemini and, since the public SynthID Detector opened,
+at [synthid.com](https://synthid.com/). That site is the `synthid-web` oracle
+surface in [`provider-oracles.md`](provider-oracles.md). On 2026-10-17 its FAQ
+named Google DeepMind, Nvidia, OpenAI, and Kakao, and its settled labels were
+`Watermark detected` and `Watermark not detected`. The old limited portal does
 not work: on 2026-09-07, its official `labs.google/synthid` entry point
 redirected a signed-in browser to a Google 404. On 2026-09-23 the same entry
 point redirected to `deepmind.google.com/frontiers/synthid`, which also returned
@@ -262,8 +262,8 @@ and implemented by `scripts/provider_oracles.py`.

 Proprietary watermark removal cannot be verified locally by design -- no public decoder
 exists. Each vendor has its own oracle and it covers only that vendor's content:
-the Content Provenance API first and `openai.com/verify` as an explicit fallback for
-OpenAI, the Gemini app for Google, and Microsoft's
+the Content Provenance API first and `openai.com/verify` for
+OpenAI, the Gemini app and the SynthID Detector at `synthid.com` for Google, and Microsoft's
 [Content Provenance Detection API](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/how-to/how-to-provenance-detection)
 for InvisMark. The Microsoft API reports `Watermark` and `C2PA` separately. It therefore
 needs a pixel-identical metadata-stripped control: the control must lose `C2PA` while
@@ -37,6 +37,12 @@ claim generator still falls back to the certificate issuer.
 Exact claim-generator mappings also cover `Grok Imagine` as xAI Grok Imagine
 and `Suno` as Suno. Both can carry a valid AI source claim without a separately
 readable issuer, so issuer-only attribution leaves the platform blank.
+Suno also [announced on 2026-08-06](https://suno.com/blog/building-the-future-of-music-responsibly)
+that it would adopt "new audio watermarking and fingerprinting technology" in
+the coming weeks. It did not name the technology
+([TechCrunch](https://techcrunch.com/2026/08/06/amid-legal-battles-suno-says-it-will-start-watermarking-songs/)
+asked whether it would be SynthID and got no answer), so there is no local
+detector, and metadata stripping leaves any such mark in the audio.

 Some signers expose the product in `signature_info` while retaining a generic
 claim generator. xAI Grok Imagine is recognized from that signed identity even
@@ -0,0 +1,4 @@
+{
+  "$schema": "https://glama.ai/mcp/schemas/server.json",
+  "maintainers": ["wiltodelta"]
+}
@@ -4,7 +4,7 @@
 # on it, including the ComfyUI node package. The console script below carries
 # the same weight, since users have it on PATH.
 name = "remove-ai-watermarks"
-version = "0.44.0"
+version = "0.46.0"
 description = "AI watermark remover for visible, invisible, and provenance marks in images and video"
 readme = "README.md"
 requires-python = ">=3.11,<3.15"
@@ -205,6 +205,9 @@ dev = [
    "httpx>=0.28,<1",
    # Development-only isolated browser automation for public provider oracles.
    "playwright>=1.62,<2",
+    # The official MCP client, used only to test `remove-ai-watermarks mcp`
+    # against the reference implementation. The server itself is stdlib-only.
+    "mcp>=2.3,<3",
    "pytest>=8.0.0",
    "pytest-cov>=4.1.0",
    "pytest-xdist>=3.5.0",
@@ -2,7 +2,7 @@

 from __future__ import annotations

-import datetime
+from datetime import date
 from typing import TYPE_CHECKING

 if TYPE_CHECKING:
@@ -10,15 +10,15 @@ if TYPE_CHECKING:
    from pathlib import Path


-def corpus_day(path: Path, corpus: Path) -> datetime.date | None:
+def corpus_day(path: Path, corpus: Path) -> date | None:
    """Read the leading YYYY-MM-DD corpus segment, if present."""
    try:
-        return datetime.date.fromisoformat(path.relative_to(corpus).parts[0])
+        return date.fromisoformat(path.relative_to(corpus).parts[0])
    except (ValueError, IndexError):
        return None


-def corpus_files(corpus: Path, since: datetime.date | None, *, suffixes: Collection[str] | None = None) -> list[Path]:
+def corpus_files(corpus: Path, since: date | None, *, suffixes: Collection[str] | None = None) -> list[Path]:
    """Return a deterministic recursive walk, optionally date- and suffix-bounded."""
    if since is None:
        candidates = corpus.rglob("*")
@@ -23,7 +23,6 @@ from the provider's canonical name, the provider is named separately.
 | `samsung_en_alpha.png` | Samsung Galaxy AI label, English locale ("AI-generated content", bottom-left) | 5 corpus files | POS 0.11-0.30 vs NEG max 0.40 (binary); tophat/gray tried, no separation. The registered Italian engine scores 0.18-0.31 on the same files -- same layout class, wrong glyph template. |
 | `gemini_text_alpha.png` | "Generated with Gemini" text label (bottom-right; the registered gemini mark is the sparkle icon) | 3 corpus files | POS 0.07-0.22 vs NEG max 0.32; coverage gate finds the blob (0.29-0.35) but the Arial silhouette misses Google's letterforms. |
 | `notebooklm_alpha.png` | NotebookLM wordmark (bottom-right) | 12 corpus files | locate geometry not yet fitted; POS max 0.12. |
-| `dola_alpha.png` | DolaAI on images (the video mark is registered) | 12 corpus files | POS 0.11-0.21 vs NEG max 0.30. |
 | `mindvideo_alpha.png` | MindVideo.AI (top-right) | 11 corpus files | POS 0.29-0.32 vs NEG max 0.30 -- borderline overlap, not shippable. |
 | `higgsfield_alpha.png` | `HIGGSFIELD AI` wordmark (Higgsfield AI, bottom-right; the boxed `AI` variant shares the cohort) | 5 wordmark files (16 in the boxed-AI OCR cluster) | POS max 0.26 vs NEG max 0.22 -- no separation; the mark may be two-part (wordmark + boxed AI) and needs a composed template. |
 | `jianying_alpha.png` | `剪映AI` wordmark (Jianying / 剪映, the China product in the CapCut family; bottom-right) | 2 corpus files | POS 0.29 vs NEG max 0.35. |
@@ -121,7 +121,7 @@ def load_pinned_models() -> tuple[object, object]:
        raw = convert_state_dict_for_scriptable_model(raw)
        model = create(config, device="cpu")  # type: ignore[call-arg]
        load_state_dict(model, state_dict=raw)
-        model.eval()
+        model.train(False)
        return model

    generator = load(GENERATOR_CARD, "generator_base.pth", AudioSealWMConfig, create_generator)
@@ -51,7 +51,7 @@ def export_onnx(checkpoint: Path, output: Path) -> None:
    if output.exists():
        raise SystemExit(f"refusing to overwrite existing output: {output}")
    output.parent.mkdir(parents=True, exist_ok=True)
-    model = VisionEmbedding(load_headed_clip(checkpoint, torch.device("cpu"))).eval()
+    model = VisionEmbedding(load_headed_clip(checkpoint, torch.device("cpu"))).train(False)
    export_options: dict[str, Any] = {}
    if "dynamo" in inspect.signature(torch.onnx.export).parameters:
        export_options["dynamo"] = False
@@ -252,7 +252,7 @@ def _lpips_model() -> tuple[Any, Any]:
    import torch

    model = lpips.LPIPS(net="alex", verbose=False)
-    model.eval()
+    model.train(False)
    return model, torch


@@ -173,7 +173,7 @@ def load_model() -> object:
        raise RuntimeError(
            f"PixelSeal checkpoint/source mismatch: missing={loaded.missing_keys}, unexpected={loaded.unexpected_keys}"
        )
-    model.eval()
+    model.train(False)
    return model


@@ -288,8 +288,8 @@ def run_web_batch(
    oracles.verify_batch(manifest_path)
    manifest, results = oracles.load_batch(manifest_path)
    surface = manifest["surface"]
-    if surface == "gemini-web":
-        raise ValueError("gemini-web requires the real Chrome workflow, not isolated Playwright")
+    if oracles.SURFACES[surface].automation == "real_browser":
+        raise ValueError(f"{surface} requires the real Chrome workflow, not isolated Playwright")
    if surface not in PLAYWRIGHT_SURFACES:
        raise ValueError(f"surface {surface} is not a Playwright Web oracle")
    slot = manifest["slot"]
@@ -169,6 +169,26 @@ SURFACES: dict[str, OracleSurface] = {
        account_slots=True,
        network_slots=True,
    ),
+    "synthid-web": OracleSurface(
+        key="synthid-web",
+        provider_key="google",
+        provider="Google",
+        signal_family="synthid",
+        surface="SynthID Detector",
+        automation="real_browser",
+        url="https://synthid.com/",
+        media_types=("image", "video", "audio"),
+        project_evidence=(),
+        prompt=None,
+        clean_result_note=(
+            "Record detected only for the explicit label Watermark detected, and not_detected only for "
+            "Watermark not detected. The portal reports partner SynthID from Google DeepMind, Nvidia, "
+            "OpenAI, and Kakao without naming which partner. A not_detected result is not proof of "
+            "human authorship. Preserve quota, sign-in, and unclear wording."
+        ),
+        account_slots=True,
+        network_slots=True,
+    ),
    "meta-web": OracleSurface(
        key="meta-web",
        provider_key="meta",
@@ -278,6 +298,11 @@ def ordered_surfaces() -> list[OracleSurface]:
    return sorted(SURFACES.values(), key=_surface_order)


+def _automation_keys(automation: str) -> frozenset[str]:
+    """Return the surface keys that use one automation kind."""
+    return frozenset(key for key, surface in SURFACES.items() if surface.automation == automation)
+
+
 def _inside(path: Path, parent: Path) -> bool:
    """Return whether resolved PATH is inside resolved PARENT."""
    try:
@@ -364,6 +389,8 @@ def load_slots(path: Path) -> dict[str, ExecutionSlot]:
        "api_key_env",
    }
    slots: dict[str, ExecutionSlot] = {}
+    real_browser = _automation_keys("real_browser")
+    playwright = _automation_keys("playwright")
    for index, raw_slot in enumerate(slot_items):
        if not isinstance(raw_slot, dict):
            raise ValueError(f"slot {index} must be a JSON object")
@@ -407,19 +434,17 @@ def load_slots(path: Path) -> dict[str, ExecutionSlot]:
            raise ValueError(f"surface {surface_key} requires an account_label")
        if account_label is not None and not SURFACES[surface_key].account_slots:
            raise ValueError(f"surface {surface_key} does not use account slots")
-        if surface_key == "gemini-web" and browser_profile is None:
-            raise ValueError("surface gemini-web requires a browser_profile")
+        if surface_key in real_browser and browser_profile is None:
+            raise ValueError(f"surface {surface_key} requires a browser_profile")
        if surface_key == "gemini-web" and google_account_index is None:
            raise ValueError("surface gemini-web requires a google_account_index")
        if surface_key != "gemini-web" and google_account_index is not None:
            raise ValueError("only gemini-web uses google_account_index")
        if SURFACES[surface_key].automation == "api" and browser_profile is not None:
            raise ValueError(f"API surface {surface_key} does not use a browser_profile")
-        if surface_key != "gemini-web" and browser_profile is not None:
-            raise ValueError("only gemini-web uses a real browser_profile")
-        if surface_key == "gemini-web" and proxy_url_env is not None:
-            raise ValueError("surface gemini-web does not use proxy_url_env")
-        if proxy_url_env is not None and surface_key not in {"openai-web", "microsoft-web", "meta-web"}:
+        if surface_key not in real_browser and browser_profile is not None:
+            raise ValueError("only a real-browser surface uses a browser_profile")
+        if proxy_url_env is not None and surface_key not in playwright:
            raise ValueError(f"surface {surface_key} does not use proxy_url_env")
        if surface_key == "openai-api" and api_key_env is None:
            raise ValueError("surface openai-api requires an api_key_env")
@@ -858,18 +883,17 @@ def list_oracles(as_json: bool) -> None:
 def plan(provider: str, as_json: bool) -> None:
    """Show the API-first surface order for one provider."""
    definitions = provider_plan(provider)
+    has_api = any(definition.automation == "api" for definition in definitions)
    rows = []
-    for index, definition in enumerate(definitions):
+    for definition in definitions:
        row = definition.to_dict()
-        row["preference"] = "primary" if index == 0 else "fallback"
+        row["preference"] = "fallback" if has_api and definition.automation != "api" else "primary"
        rows.append(row)
    if as_json:
        click.echo(json.dumps(rows, indent=2, sort_keys=True))
        return
    for row in rows:
        click.echo(f"{row['preference']}\t{row['key']}\t{row['automation']}\t{row['surface']}")
-    if len(rows) > 1:
-        click.echo("Web fallback requires a separate operator decision; failures do not trigger it automatically.")


@cli.command()
@@ -0,0 +1,68 @@
+"""Rebuild synthetic silhouettes for additional AI disclosure layouts."""
+
+from __future__ import annotations
+
+import logging
+import sys
+from pathlib import Path
+
+import cv2
+import numpy as np
+from PIL import Image
+
+ROOT = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(ROOT))
+from scripts.render_vendor_silhouettes import render  # noqa: E402
+
+log = logging.getLogger(__name__)
+ASSETS = ROOT / "src" / "remove_ai_watermarks" / "assets"
+CJK = {"font": "/System/Library/Fonts/Hiragino Sans GB.ttc", "font_index": 2}
+
+
+def centered_lines(
+    top: str, bottom: str, *, top_scale: float, bottom_height: float = 1.0, shear: float = 0.0
+) -> np.ndarray:
+    upper, lower = render(top, opts=CJK), render(bottom, opts=CJK)
+    if bottom_height != 1.0:
+        lower = cv2.resize(lower, (upper.shape[1], round(upper.shape[0] * bottom_height)))
+    else:
+        upper = cv2.resize(upper, (round(lower.shape[1] * top_scale), round(lower.shape[0] * 0.95)))
+    lines = [upper, lower]
+    result = np.zeros((sum(line.shape[0] for line in lines) + 2, max(line.shape[1] for line in lines)), np.uint8)
+    y = 0
+    for line in lines:
+        x = (result.shape[1] - line.shape[1]) // 2
+        result[y : y + line.shape[0], x : x + line.shape[1]] = line
+        y += line.shape[0] + 2
+    if shear:
+        h, w = result.shape
+        extra = round(abs(shear) * h)
+        result = cv2.warpAffine(result, np.float32([[1, shear, extra], [0, 1, 0]]), (w + extra, h))
+    return result
+
+
+def silhouettes() -> dict[str, np.ndarray]:
+    return {
+        "workbuddy_alpha.png": centered_lines("AI生成", "WORKBUDDY>_", top_scale=1.0, bottom_height=0.45),
+        "yuanbao_compact_alpha.png": centered_lines("元宝", "AI生成", top_scale=0.65, shear=-0.15),
+        "yuanbao_compact_gray_alpha.png": centered_lines("元宝", "AI生成", top_scale=0.65, shear=-0.15),
+        "doubao_outline_alpha.png": render("豆包AI生成", opts=CJK),
+        "samsung_ko_alpha.png": render(
+            "AI로 생성한 콘텐츠", opts={"font": "/System/Library/Fonts/AppleSDGothicNeo.ttc"}
+        ),
+        "microsoft_text_alpha.png": render(
+            "Made with AI", opts={"font": "/System/Library/Fonts/Supplemental/Arial.ttf"}
+        ),
+        "jimeng_label_alpha.png": render("AI生成", opts={"font": "/System/Library/Fonts/STHeiti Medium.ttc"}),
+    }
+
+
+def main() -> None:
+    logging.basicConfig(level=logging.INFO)
+    for name, alpha in silhouettes().items():
+        Image.fromarray(alpha).save(ASSETS / name)
+        log.info("Rendered %s, %s x %s", name, alpha.shape[1], alpha.shape[0])
+
+
+if __name__ == "__main__":
+    main()
@@ -73,7 +73,7 @@ MARKS = {
    "gemini_text_alpha.png": "Generated with Gemini",
    # Candidate wordmarks measured on a local evaluation corpus (unregistered).
    "notebooklm_alpha.png": "NotebookLM",
-    "dola_alpha.png": "DolaAI",
+    "dola_alpha.png": "Dola AI",
    "mindvideo_alpha.png": "MindVideo.AI",
    "higgsfield_alpha.png": "HIGGSFIELD AI",
    "capcut_alpha.png": "CapCut AI",
@@ -92,6 +92,7 @@ _REGISTERED = {f"{key}_alpha.png" for key in mark_keys()} & MARKS.keys()

 # Per-mark post-processing for the multi-line / slanted stamps (see render()).
 MARK_OPTS: dict[str, dict[str, Any]] = {
+    "dola_alpha.png": {"font": "/System/Library/Fonts/Supplemental/Arial Bold.ttf"},
    # Hiragino Sans GB W6, tight leading, dilation, and negative shear match the
    # standard Yuanbao stamp without clipping the lower line.
    "yuanbao_alpha.png": {
@@ -98,7 +98,7 @@ def _train_detector(emb: dict[str, np.ndarray], catalog: dict) -> tuple[nn.Modul
            optimizer.zero_grad()
            loss_fn(model(features).squeeze(-1), labels).backward()
            optimizer.step()
-    model.eval()
+    model.train(False)
    dev = np.stack(
        [
            emb[row["sha256"]]
@@ -161,7 +161,7 @@ def _train_provider(feat: dict[str, np.ndarray], catalog: dict) -> dict[str, nn.
                optimizer.zero_grad()
                _focal_loss(model(features).squeeze(-1), labels).backward()
                optimizer.step()
-        heads[name] = model.eval()
+        heads[name] = model.train(False)
        log.info("provider %s pos=%d", name, len(positive))
    return heads

@@ -127,7 +127,7 @@ def load_model() -> object:
    import torch

    model = torch.jit.load(str(checkpoint_path()), map_location="cpu")
-    model.eval()
+    model.train(False)
    return model


@@ -0,0 +1,30 @@
+{
+  "$schema": "https://static.modelcontextprotocol.io/schemas/2025-12-11/server.schema.json",
+  "name": "io.github.wiltodelta/remove-ai-watermarks",
+  "title": "Remove AI Watermarks",
+  "description": "Identify and remove AI watermarks and provenance metadata from your own images and video.",
+  "version": "0.46.0",
+  "repository": {
+    "url": "https://github.com/wiltodelta/remove-ai-watermarks",
+    "source": "github"
+  },
+  "websiteUrl": "https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/mcp-server.md",
+  "packages": [
+    {
+      "registryType": "pypi",
+      "registryBaseUrl": "https://pypi.org",
+      "identifier": "remove-ai-watermarks",
+      "version": "0.46.0",
+      "runtimeHint": "uvx",
+      "transport": {
+        "type": "stdio"
+      },
+      "packageArguments": [
+        {
+          "type": "positional",
+          "value": "mcp"
+        }
+      ]
+    }
+  ]
+}
@@ -1,7 +1,7 @@
 {
  "name": "remove-ai-watermarks",
  "displayName": "Remove AI Watermarks",
-  "version": "1.0.10",
+  "version": "1.0.12",
  "description": "Identify and remove AI provenance marks from images and video the user generated or edited, using the remove-ai-watermarks CLI.",
  "author": {
    "name": "wiltodelta",
@@ -5,7 +5,7 @@ license: Apache-2.0
 compatibility: Python 3.11-3.14. Invisible image removal needs NVIDIA CUDA. Some video writes need ffmpeg. Installer may be uv, pipx, or pip.
 metadata:
  author: wiltodelta
-  version: "1.0.10"
+  version: "1.0.12"
  homepage: https://raiw.cc
  repository: https://github.com/wiltodelta/remove-ai-watermarks
 ---
@@ -21,8 +21,8 @@ remove-ai-watermarks visible image.png --keep-metadata -o clean.png
 ```

 Image mark keys: `gemini`, `doubao`, `jimeng`, `qwen`, `wan`, `kling`, `yuanbao`,
-`samsung`, `runninghub`, `baidu`, `liblib`, `liblib_pill`, `microsoft`,
-`generic_ai_label`, `openart`, `jimeng_pill`. `wan` is new (since CLI 0.43.0). Default
+`dola`, `workbuddy`, `samsung`, `runninghub`, `baidu`, `liblib`, `liblib_pill`, `microsoft`,
+`generic_ai_label`, `openart`, `jimeng_pill`. `dola` and `workbuddy` are new (since CLI 0.45.0). Default
 `--mark auto` removes every selected match. `microsoft` is the top-right AI
 badge, separate from the invisible InvisMark watermark on the same vendor's
 images. `jimeng_pill` is a weak detector and needs corroboration. The compact
@@ -138,6 +138,13 @@ user names a device, and never to claim a device the probe did not report.
 Without `-o`, video commands write `<source>_clean` and keep the original.
 Output extension must match the input container.

+## MCP server
+
+`remove-ai-watermarks mcp` (since CLI 0.46.0) serves `identify`, `remove`
+and `batch` as MCP tools over stdin and stdout for MCP clients. It blocks
+waiting for protocol messages, so do not run it from this skill; call the
+commands above directly.
+
 ## Exit codes

 | Code | Meaning |
@@ -32,7 +32,7 @@ _GPU_MARKERS = ("GPU ", "gpu:")
 # told an agent its CLI was current right before it typed a value that build
 # rejects. tests/test_agent_skill.py pins it against the package version and
 # against every "(since CLI X.Y.Z)" annotation in the skill.
-MIN_CLI_VERSION = (0, 43, 0)
+MIN_CLI_VERSION = (0, 46, 0)
 # Markers a build without the pixel stack prints. The first is the current CLI's own
 # install hint; the rest are what older builds emit before the guard existed, and the
 # Homebrew formula ships exactly such a build.
@@ -35,7 +35,7 @@ _os.environ.setdefault("TRANSFORMERS_VERBOSITY", "error")
 _warnings.filterwarnings("ignore", message=r".*ImageProcessorFast.*")


-__version__ = "0.44.0"
+__version__ = "0.46.0"

 __all__ = [
    "BatchItemResult",
@@ -1,7 +1,7 @@
 """Generated C2PA soft-binding registry; do not edit by hand.

 Source: https://raw.githubusercontent.com/c2pa-org/softbinding-algorithm-list/main/softbinding-algorithm-list.json
-Revision: d36d1c1742b1753ca064bb3307abfd95db912a28
+Revision: d5960248ab69c17dc194906d2ef7d11ad164be70
 Source license: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
 Changes: validated, ordered, and reduced to fields used at runtime; descriptions
 are shortened to their first sentence for display.
@@ -12,7 +12,7 @@ from __future__ import annotations
 C2PA_SOFT_BINDING_SOURCE_URL = (
    "https://raw.githubusercontent.com/c2pa-org/softbinding-algorithm-list/main/softbinding-algorithm-list.json"
 )
-C2PA_SOFT_BINDING_SOURCE_REVISION = "d36d1c1742b1753ca064bb3307abfd95db912a28"
+C2PA_SOFT_BINDING_SOURCE_REVISION = "d5960248ab69c17dc194906d2ef7d11ad164be70"
 C2PA_SOFT_BINDING_SOURCE_LICENSE = "CC BY 4.0"

 C2PA_SOFT_BINDING_ROWS: tuple[
@@ -632,4 +632,15 @@ C2PA_SOFT_BINDING_ROWS: tuple[
        (),
        False,
    ),
+    (
+        54,
+        "be.immostory.videoseal.1",
+        "watermark",
+        ("video",),
+        ("video/mp4",),
+        "ImmoStory AI deployment of Meta FAIR's Video Seal 1.0 invisible video watermark",
+        "2026-10-12T00:00:00.000Z",
+        (),
+        False,
+    ),
 )
@@ -67,7 +67,7 @@ def load_headed_clip(checkpoint: Path, device: torch.device) -> HeadedCLIP:
        model = HeadedCLIP(clip)
    state = torch.load(checkpoint, map_location="cpu", weights_only=True)
    model.load_state_dict(state, strict=True)
-    return model.to(device).eval()
+    return model.to(device).train(False)


 def load_processor() -> Any:
@@ -7,7 +7,9 @@ without re-encoding, the image path drops matching boxes and emits the rest
 verbatim. The streaming MP4/MOV/M4A path instead preserves all offsets by
 retyping matching boxes as ``free`` and blanking their payloads in place. The
 codestream (``mdat`` for ISOBMFF, ``jxlc`` / ``jxlp`` for JPEG-XL) is untouched,
-so pixel, video, and audio data is preserved bit-for-bit.
+so pixel, video, and audio data is preserved bit-for-bit. The one exception is
+Kling's SEI tag inside ``mdat``: its bytes are blanked in place, which no decoder
+reads, so every decoded frame stays identical.

 TC260-PG-20257A video metadata is nested instead:
 ``moov.udta.meta.keys/ilst``. Its detector seeks through those boxes without
@@ -43,8 +45,11 @@ from remove_ai_watermarks.metadata import (
    C2PA_UUID,
    IPTC_AI_FIELD_MARKERS,
    IPTC_AI_MARKERS,
+    KLING_SEI_TAG,
    LOCAL_GENERATION_METADATA_KEYS,
    MAX_TC260_VALUE_BYTES,
+    app_generator_from_metadata,
+    app_provenance_from_metadata,
    parse_tc260_aigc_json,
 )

@@ -347,6 +352,55 @@ def _ai_metadata_regions(
    return regions


+def _app_metadata_regions(
+    stream: BinaryIO,
+    file_size: int,
+) -> list[tuple[tuple[int, int] | None, int, int]]:
+    """Locate keyed or keyless values carrying an app-export provenance object.
+
+    Dreamina writes its export JSON under ordinary keys (``LvMetaInfo`` and a
+    second copy under ``com.apple.quicktime.artwork``). The file-level scan reads
+    it from the raw head bytes, so removal must blank every item it would match.
+    """
+    return [
+        (key_span, value_start, value_end)
+        for _name, key_span, value_start, value_end, value in _metadata_items(stream, file_size)
+        if app_provenance_from_metadata(value) or app_generator_from_metadata(value)
+    ]
+
+
+def _copy_finding_kling_sei(
+    source: BinaryIO,
+    target: BinaryIO,
+    mdat_ranges: list[tuple[int, int]],
+) -> list[int]:
+    """Copy ``source`` to ``target`` and return each Kling SEI tag offset in ``mdat``.
+
+    Detection reads the tag from the head window, but removal covers every copy in
+    the media data. The search rides on the copy, so the file is read once. Each
+    window carries one byte less than the tag from the previous chunk, so a copy
+    straddling a read boundary is found exactly once.
+    """
+    keep = len(KLING_SEI_TAG) - 1
+    offsets: list[int] = []
+    tail = b""
+    pos = 0
+    while chunk := source.read(_STREAM_COPY_BYTES):
+        target.write(chunk)
+        if mdat_ranges:
+            window = tail + chunk
+            base = pos - len(tail)
+            found = window.find(KLING_SEI_TAG)
+            while found != -1:
+                start = base + found
+                if any(low <= start and start + len(KLING_SEI_TAG) <= high for low, high in mdat_ranges):
+                    offsets.append(start)
+                found = window.find(KLING_SEI_TAG, found + 1)
+            tail = window[-keep:]
+        pos += len(chunk)
+    return offsets
+
+
 def ai_metadata_tags(path: str | Path, *, strict: bool = False) -> dict[str, str]:
    """Read keyed AI-generation metadata from an ISOBMFF container."""
    try:
@@ -601,15 +655,16 @@ def strip_isobmff_media_file(

    The output retains every box size and byte offset. A top-level C2PA/JUMBF or
    AI-label box is converted to a ``free`` box and its payload is zeroed; native
-    TC260 key/value spans are blanked in place. Keeping the original lengths is
-    required because removing a pre-``mdat`` box would invalidate absolute media
-    offsets in an existing sample table.
+    TC260, generation-key and app-export key/value spans, and the Kling SEI tag in
+    ``mdat``, are blanked in place. Keeping the original lengths is required
+    because removing a pre-``mdat`` box would invalidate absolute media offsets in
+    an existing sample table.

    The source is copied in bounded chunks to a sibling temporary file and
    atomically published only after all patches succeed. A malformed top-level
    walk is fail-safe: the input is copied unchanged.

-    Returns ``(provenance_boxes_blanked, native_tc260_keys_blanked)``.
+    Returns ``(provenance_boxes_blanked, native_tags_blanked)``.
    """
    from pathlib import Path as _Path

@@ -627,12 +682,23 @@ def strip_isobmff_media_file(
        )
        tc260_regions = _tc260_aigc_regions(stream, file_size) if targets is not None else []
        ai_metadata_regions = _ai_metadata_regions(stream, file_size) if targets is not None else []
+        app_regions = _app_metadata_regions(stream, file_size) if targets is not None else []
+        mdat_ranges = (
+            [
+                (payload, end)
+                for _start, end, box_type, payload in iter_file_boxes(stream, 0, file_size)
+                if box_type == b"mdat"
+            ]
+            if targets is not None
+            else []
+        )
        tc260_key_spans = {region[0] for region in tc260_regions if region[0] is not None}
        ai_metadata_key_spans = {region[0] for region in ai_metadata_regions}
+        app_key_spans = {region[0] for region in app_regions if region[0] is not None}

    with atomic_video_output(output_path) as temporary_path:
        with source_path.open("rb") as source_stream, temporary_path.open("r+b") as temporary:
-            shutil.copyfileobj(source_stream, temporary, length=_STREAM_COPY_BYTES)
+            kling_sei_offsets = _copy_finding_kling_sei(source_stream, temporary, mdat_ranges)
            if targets is not None:
                for box_start, payload_start, box_end in targets:
                    temporary.seek(box_start + 4)
@@ -647,6 +713,15 @@ def strip_isobmff_media_file(
                    temporary.seek(key_span[0])
                    temporary.write(b"free" + b" " * (key_span[1] - key_span[0] - 4))
                    _overwrite_range(temporary, value_start, value_end, byte=b" ")
+                for key_span, value_start, value_end in app_regions:
+                    if key_span is not None:
+                        temporary.seek(key_span[0])
+                        temporary.write(b"free" + b" " * (key_span[1] - key_span[0] - 4))
+                    _overwrite_range(temporary, value_start, value_end, byte=b" ")
+                # Spaces keep the SEI message well-formed: its length fields stand,
+                # and a 0x20 run cannot form an emulation-prevention sequence.
+                for offset in kling_sei_offsets:
+                    _overwrite_range(temporary, offset, offset + len(KLING_SEI_TAG), byte=b" ")
            temporary.flush()
            os.fsync(temporary.fileno())
        shutil.copymode(source_path, temporary_path)
@@ -657,7 +732,8 @@ def strip_isobmff_media_file(
            source_path,
        )
        return 0, 0
-    return len(targets), len(tc260_key_spans | ai_metadata_key_spans)
+    # Every copy of the Kling tag is one tag, however many samples repeat it.
+    return len(targets), len(tc260_key_spans | ai_metadata_key_spans | app_key_spans) + bool(kling_sei_offsets)


 def strip_c2pa_boxes(data: bytes) -> tuple[bytes, int]:
@@ -47,6 +47,12 @@ _LATENT_GRID = 8
 # The reference graph loads the Qwen LoRA at 0.8; carrying that number to a different
 # LoRA on a different architecture would be imitation, not parity.
 SDXL_STEPS = 4
+# Below this floor the global stack streams its weights from CPU layer by layer; at or
+# above it the stack stays resident. The fp16 UNet, Canny ControlNet, both text
+# encoders and the VAE are ~8.8 GiB of weights, and the UNet and the ControlNet run
+# together on every step, so model-level offload cannot get under an 8 GiB card.
+# Benchmark in docs/module-internals.md, "CPU offload".
+RESIDENT_SDXL_GLOBAL_MIN_VRAM_GIB = 12.0


 def sdxl_target_size(width: int, height: int) -> tuple[int, int]:
@@ -54,6 +60,17 @@ def sdxl_target_size(width: int, height: int) -> tuple[int, int]:
    return _target_size(width, height, _LATENT_GRID)


+def resolve_sdxl_global_residency(
+    requested: bool | None,
+    *,
+    total_memory_gib: float,
+) -> bool:
+    """Keep the SDXL global stack resident when explicitly requested or safely sized."""
+    if requested is not None:
+        return requested
+    return total_memory_gib >= RESIDENT_SDXL_GLOBAL_MIN_VRAM_GIB
+
+
@dataclass
 class SdxlZImagePipeline(TwoStageZImagePipeline):
    """Lazy runtime for the SDXL global stage plus the inherited face stage."""
@@ -64,6 +81,12 @@ class SdxlZImagePipeline(TwoStageZImagePipeline):
        super().__post_init__()
        self._sdxl_pipe: Any = None

+    def _keep_global_models_resident(self) -> bool:
+        return resolve_sdxl_global_residency(
+            self.keep_global_models_on_device,
+            total_memory_gib=self._total_vram_gib(),
+        )
+
    def _load_global(self) -> Any:
        if self._sdxl_pipe is not None:
            return self._sdxl_pipe
@@ -89,11 +112,19 @@ class SdxlZImagePipeline(TwoStageZImagePipeline):
            variant="fp16",
            add_watermarker=False,
            **token,
-        ).to(self.device)
+        )
+        resident = self._keep_global_models_resident()
+        if resident:
+            pipe = pipe.to(self.device)
        pipe.load_lora_weights(hf_hub_download(SDXL_LIGHTNING_MODEL_ID, SDXL_LIGHTNING_PATTERN, **token))
        pipe.fuse_lora()
        # SDXL-Lightning is distilled against trailing timestep spacing.
        pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing")
+        if not resident:
+            # After the fuse: the offload hooks keep their own CPU copy of each weight
+            # and stream that copy, so fusing later would never reach what runs.
+            self._progress("Streaming the SDXL stack from CPU (card below the residency floor)...")
+            pipe.enable_sequential_cpu_offload(device=self.device, gpu_id=torch.cuda.current_device())
        self._sdxl_pipe = pipe
        return pipe

@@ -823,7 +823,7 @@ class TwoStageZImagePipeline:
            torch_dtype=self._face_stage_dtype(),
            **kwargs,
        ).to(self.device)
-        model.eval()
+        model.train(False)
        self._sam_model = model
        self._sam_processor = processor
        return model, processor
@@ -63,13 +63,13 @@ PROFILE_ADAPTIVE_POLISH = {

 # Which profiles let ``--cpu-offload`` reach their GLOBAL stack. The face stage
 # honours the flag everywhere, because residency lives on the shared base, but the
-# global stack is each profile's own: only qwen-zimage streams it, through
-# ``_qwen_vram_config``. chroma-zimage and sdxl-zimage load theirs with a plain
-# ``.to(device)``, so on a card below the face-stage residency floor the flag has
-# nothing left to change and the run behaves as if it were absent. Data rather than
-# a subclass attribute, because the caller that must warn cannot import a profile
-# module without pulling in Diffusers.
-GLOBAL_OFFLOAD_PROFILES = frozenset({QWEN_ZIMAGE_PROFILE})
+# global stack is each profile's own: qwen-zimage streams it through
+# ``_qwen_vram_config`` and sdxl-zimage through Diffusers sequential offload.
+# chroma-zimage loads its stack with a plain ``.to(device)``, so on a card below the
+# face-stage residency floor the flag has nothing left to change there and the run
+# behaves as if it were absent. Data rather than a subclass attribute, because the
+# caller that must warn cannot import a profile module without pulling in Diffusers.
+GLOBAL_OFFLOAD_PROFILES = frozenset({QWEN_ZIMAGE_PROFILE, SDXL_ZIMAGE_PROFILE})

 SDXL_LIGHTNING_MODEL_ID = "ByteDance/SDXL-Lightning"
 SDXL_LIGHTNING_PATTERN = "sdxl_lightning_4step_lora.safetensors"
@@ -175,9 +175,9 @@ class WatermarkRemover:
            return
        logger.warning(
            "--cpu-offload does not reach the global stack of the '%s' pipeline: only "
-            "qwen-zimage streams its global model between CUDA calls. The face stage "
-            "still honours the flag, so below the face-stage residency floor this run "
-            "behaves as if --cpu-offload were absent.",
+            "qwen-zimage and sdxl-zimage stream their global model between CUDA calls. "
+            "The face stage still honours the flag, so below the face-stage residency "
+            "floor this run behaves as if --cpu-offload were absent.",
            self.model_profile,
        )

@@ -181,6 +181,7 @@ class TextMarkDetection:
    # ``footprint_mask`` reuses a threaded detection only when it matches the STRICT
    # level its own re-detect would have used -- see TextMarkEngine._strict_detection.
    provenance: bool = False
+    template_asset: str | None = None


 def best_detection(*detections: TextMarkDetection) -> TextMarkDetection:
@@ -665,7 +666,7 @@ class TextMarkEngine:
    def _verdict(self, scan: TextMarkScan, *, provenance: bool) -> TextMarkDetection:
        """Apply the trust-level-dependent tail to a scan, as a fresh result object."""
        c = self.config
-        det = TextMarkDetection(provenance=provenance)
+        det = TextMarkDetection(provenance=provenance, template_asset=c.asset_name)
        if scan.loc is None or scan.box is None:
            return det
        det.region = scan.loc.bbox
@@ -0,0 +1,219 @@
+"""Combine independently gated text layouts and retain their mask geometry."""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Any
+
+import cv2
+
+from remove_ai_watermarks._text_mark_engine import (
+    BottomRightAnchoredEngine,
+    TextMarkConfig,
+    TextMarkDetection,
+    TextMarkEngine,
+    TextMarkLocation,
+    best_detection,
+)
+
+if TYPE_CHECKING:
+    from numpy.typing import NDArray
+
+
+class TextMarkVariantsMixin(TextMarkEngine):
+    """Add independently configured layouts to an existing text-mark engine."""
+
+    _variants: tuple[TextMarkEngine, ...] = ()
+    config: TextMarkConfig
+
+    def detect_both(self, image: NDArray[Any] | None) -> tuple[TextMarkDetection, TextMarkDetection]:
+        pairs = [super().detect_both(image), *(engine.detect_both(image) for engine in self._variants)]
+        return best_detection(*(pair[0] for pair in pairs)), best_detection(*(pair[1] for pair in pairs))
+
+    def detect(self, image: NDArray[Any], *, provenance: bool = False) -> TextMarkDetection:
+        return self.detect_both(image)[int(provenance)]
+
+    def variant_engine(self, detection: TextMarkDetection) -> TextMarkEngine:
+        for engine in self._variants:
+            if engine.config.asset_name == detection.template_asset:
+                return engine
+        return self
+
+    def footprint_mask(
+        self,
+        image: NDArray[Any] | None,
+        *,
+        force: bool = False,
+        dilate: int | None = None,
+        detection: TextMarkDetection | None = None,
+    ) -> NDArray[Any] | None:
+        if image is None or image.size == 0 or force:
+            return super().footprint_mask(image, force=force, dilate=dilate, detection=detection)
+        det = detection if detection is not None else self.detect(image)
+        engine = self.variant_engine(det)
+        if engine is self:
+            return super().footprint_mask(image, force=False, dilate=dilate, detection=det)
+        return engine.footprint_mask(image, force=False, dilate=dilate, detection=det)
+
+
+class GlyphTextMarkEngine(TextMarkEngine):
+    """Use the accepted template's aligned strokes instead of a textured ROI."""
+
+    def footprint_mask(
+        self,
+        image: NDArray[Any] | None,
+        *,
+        force: bool = False,
+        dilate: int | None = None,
+        detection: TextMarkDetection | None = None,
+    ) -> NDArray[Any] | None:
+        from remove_ai_watermarks._text_mark_engine import load_alpha_template
+
+        if force:
+            return super().footprint_mask(image, force=True, dilate=dilate, detection=detection)
+        if image is None or image.size == 0:
+            return None
+        det = detection if detection is not None else self.detect(image)
+        alpha = load_alpha_template(self.config.asset_name)
+        if alpha is None or not det.detected:
+            return None
+        return self._aligned_alpha_mask(
+            image,
+            det,
+            alpha,
+            alpha_floor=0.05,
+            dilate=2 if dilate is None else max(0, dilate),
+        )
+
+
+class BottomRightGlyphTextMarkEngine(GlyphTextMarkEngine, BottomRightAnchoredEngine):
+    """Glyph-mask variant whose accepted match must hug the bottom-right corner."""
+
+    _ANCHOR_MAX_RIGHT = 0.04
+    _ANCHOR_MAX_BOTTOM = 0.04
+
+    def _ladder_best(
+        self, image: NDArray[Any], loc: TextMarkLocation
+    ) -> tuple[float, tuple[int, int, int, int] | None]:
+        """Find the strongest match among placements that satisfy the anchor."""
+        response = self._detect_response(image, loc)
+        silhouette = self._glyph_silhouette()
+        if response is None or silhouette is None:
+            return 0.0, None
+        frame_h, frame_w = image.shape[:2]
+        anchor_base = min(frame_h, frame_w)
+        base = self.scale_base(image)
+        best_score = 0.0
+        best_box: tuple[int, int, int, int] | None = None
+        for scale in self.config.ladder:
+            glyph_w = max(self.config.min_gw, int(self.config.alpha_width_frac * base * scale))
+            glyph_h = max(4, int(self.config.alpha_height_frac * base * scale))
+            if glyph_w >= response.shape[1] or glyph_h >= response.shape[0]:
+                continue
+            template = cv2.resize(silhouette, (glyph_w, glyph_h), interpolation=cv2.INTER_AREA)
+            scores = cv2.matchTemplate(response, template, cv2.TM_CCOEFF_NORMED)
+            min_x = max(0, round(frame_w - loc.x - glyph_w - self._ANCHOR_MAX_RIGHT * anchor_base))
+            max_x = min(scores.shape[1] - 1, frame_w - loc.x - glyph_w)
+            min_y = max(0, round(frame_h - loc.y - glyph_h - self._ANCHOR_MAX_BOTTOM * anchor_base))
+            max_y = min(scores.shape[0] - 1, frame_h - loc.y - glyph_h)
+            if min_x > max_x or min_y > max_y:
+                continue
+            anchored = scores[min_y : max_y + 1, min_x : max_x + 1]
+            _, score, _, point = cv2.minMaxLoc(anchored)
+            if score > best_score:
+                x, y = min_x + int(point[0]), min_y + int(point[1])
+                best_score = float(score)
+                best_box = (x, y, x + glyph_w - 1, y + glyph_h - 1)
+        return best_score, best_box
+
+
+def _box_footprint(
+    engine: GlyphTextMarkEngine,
+    image: NDArray[Any] | None,
+    *,
+    force: bool,
+    dilate: int | None,
+    detection: TextMarkDetection | None,
+    pad_left_height: float,
+    pad_right_height: float,
+) -> NDArray[Any] | None:
+    """Mask the full winning text rectangle for outlined or stacked layouts."""
+    if image is None or image.size == 0 or force:
+        return GlyphTextMarkEngine.footprint_mask(
+            engine,
+            image,
+            force=force,
+            dilate=dilate,
+            detection=detection,
+        )
+    det = detection if detection is not None else engine.detect(image)
+    if not det.detected or det.match_box is None:
+        return None
+    from remove_ai_watermarks import region_eraser
+
+    loc = engine.locate(image)
+    x0, y0, x1, y1 = det.match_box
+    extra_left = round((y1 - y0 + 1) * pad_left_height)
+    extra_right = round((y1 - y0 + 1) * pad_right_height)
+    pad = 2 if dilate is None else max(0, dilate)
+    return region_eraser.boxes_to_mask(
+        image.shape[:2],
+        [
+            (
+                loc.x + x0 - extra_left,
+                loc.y + y0,
+                x1 - x0 + 1 + extra_left + extra_right,
+                y1 - y0 + 1,
+            )
+        ],
+        dilate=pad,
+    )
+
+
+class BoxTextMarkEngine(GlyphTextMarkEngine):
+    """Text detector with a solid detector-aligned removal footprint."""
+
+    _BOX_PAD_LEFT_HEIGHT = 0.0
+    _BOX_PAD_RIGHT_HEIGHT = 0.0
+
+    def footprint_mask(
+        self,
+        image: NDArray[Any] | None,
+        *,
+        force: bool = False,
+        dilate: int | None = None,
+        detection: TextMarkDetection | None = None,
+    ) -> NDArray[Any] | None:
+        return _box_footprint(
+            self,
+            image,
+            force=force,
+            dilate=dilate,
+            detection=detection,
+            pad_left_height=self._BOX_PAD_LEFT_HEIGHT,
+            pad_right_height=self._BOX_PAD_RIGHT_HEIGHT,
+        )
+
+
+class BottomRightBoxTextMarkEngine(BottomRightGlyphTextMarkEngine):
+    """Anchored text detector with a solid detector-aligned footprint."""
+
+    _BOX_PAD_LEFT_HEIGHT = 0.0
+    _BOX_PAD_RIGHT_HEIGHT = 0.0
+
+    def footprint_mask(
+        self,
+        image: NDArray[Any] | None,
+        *,
+        force: bool = False,
+        dilate: int | None = None,
+        detection: TextMarkDetection | None = None,
+    ) -> NDArray[Any] | None:
+        return _box_footprint(
+            self,
+            image,
+            force=force,
+            dilate=dilate,
+            detection=detection,
+            pad_left_height=self._BOX_PAD_LEFT_HEIGHT,
+            pad_right_height=self._BOX_PAD_RIGHT_HEIGHT,
+        )
@@ -467,7 +467,7 @@ def _load_runtime(device: Any, backend: ClassifierBackend = "torch") -> _Runtime
    probe = np.load(_weight_path(folder, PROBE_FILE))
    detector = _build_detector()
    detector.load_state_dict(torch.load(_weight_path(folder, DETECTOR_FILE), map_location="cpu", weights_only=True))
-    detector.to(device).eval()
+    detector.to(device).train(False)
    packed = torch.load(_weight_path(folder, PROVIDER_FILE), map_location="cpu", weights_only=True)
    heads = {}
    for name in PROVIDER_LABELS:
@@ -476,7 +476,7 @@ def _load_runtime(device: Any, backend: ClassifierBackend = "torch") -> _Runtime
            continue
        head = _build_provider_head()
        head.load_state_dict(packed[name])
-        heads[name] = head.to(device).eval()
+        heads[name] = head.to(device).train(False)
    ridge_threshold = float(probe["thr_oi_1pct"])
    mlp_threshold, provider_margin = _operating_point(folder)
    log.info("loaded photo-classify freeze from %s", folder)
@@ -308,8 +308,9 @@ _cpu_offload_option = click.option(
        "Offload model components to CPU between CUDA calls instead of keeping the "
        "whole pipeline in VRAM, at the cost of speed. Forces the face stack to "
        "offload instead of using automatic residency, on every profile. It reaches "
-        "the global stack of qwen-zimage only: chroma-zimage and sdxl-zimage keep "
-        "theirs resident either way."
+        "the global stack of qwen-zimage and sdxl-zimage, which also stream theirs "
+        "automatically on a card too small to hold it; chroma-zimage keeps its "
+        "global stack resident either way."
    ),
 )

@@ -914,6 +915,11 @@ def cmd_invisible(
    source = _validate_image(source)
    if output is None:
        output = source.with_stem(source.stem + "_clean")
+    # The scrub gate reads an undecodable file as "no target", which would exit 2
+    # (nothing to remove) for a file this command never looked at.
+    if image_io.imread(source) is None:
+        console.print(f"Error: Failed to read image: {source}")
+        raise SystemExit(1)

    # An explicit --vendor wins over detection (see the option help) and implies the
    # scrub runs: naming the cohort asserts the pixel watermark is present, so the
@@ -1905,5 +1911,19 @@ def cmd_batch(
        raise SystemExit(1)


+# ── MCP server ──
+@main.command("mcp")
+def cmd_mcp() -> None:
+    """Run a Model Context Protocol server on stdin/stdout.
+
+    Exposes identify, remove and batch as MCP tools for agent clients. Each tool
+    runs one command of this CLI, so results and exit codes match it. Start it as
+    `uvx remove-ai-watermarks mcp` from the client's server configuration.
+    """
+    from remove_ai_watermarks.mcp_server import serve
+
+    serve()
+
+
 if __name__ == "__main__":
    main()
@@ -0,0 +1,35 @@
+"""Detect the bottom-right Dola AI image disclosure, independently of video marks."""
+
+from remove_ai_watermarks._text_mark_engine import TextMarkConfig
+from remove_ai_watermarks._text_mark_variants import BoxTextMarkEngine
+
+_CONFIG = TextMarkConfig(
+    name="Dola AI image disclosure",
+    asset_name="dola_alpha.png",
+    corner="br",
+    margin_floor=4,
+    width_frac=0.28,
+    height_frac=0.09,
+    margin_x_frac=0.002,
+    margin_bottom_frac=0.002,
+    max_saturation=255,
+    logo_min_luma=110,
+    tophat_delta=8,
+    morph_open_size=3,
+    detect_min_coverage=0.01,
+    detect_ncc_threshold=0.50,
+    alpha_width_frac=0.10,
+    alpha_height_frac=0.10 * 86 / 347,
+    detect_frontend="tophat",
+    scale_basis="width",
+    ladder=(0.95, 1.0, 1.05, 1.20, 1.26),
+    provenance_ncc_factor=1.0,
+    min_gw=24,
+)
+
+
+class DolaEngine(BoxTextMarkEngine):
+    """Detect and localize the Dola AI image wordmark."""
+
+    def __init__(self) -> None:
+        super().__init__(_CONFIG)
@@ -25,10 +25,12 @@ from __future__ import annotations
 import dataclasses
 from typing import TYPE_CHECKING, Any

+import cv2
 import numpy as np

 from remove_ai_watermarks import _text_mark_engine, image_io
 from remove_ai_watermarks._text_mark_engine import TextMarkConfig, TextMarkDetection, TextMarkEngine, TextMarkScan
+from remove_ai_watermarks._text_mark_variants import BottomRightBoxTextMarkEngine, TextMarkVariantsMixin

 if TYPE_CHECKING:
    from numpy.typing import NDArray
@@ -117,6 +119,41 @@ _CONFIG = TextMarkConfig(
 _TINTED_ENGINE = TextMarkEngine(dataclasses.replace(_CONFIG, max_saturation=_TINTED_MAX_SATURATION))


+class _OutlinedDoubaoEngine(BottomRightBoxTextMarkEngine):
+    # The outlined layout's provenance-relaxed NCC floor overlaps ordinary corner
+    # texture. Require bright neutral strokes and edge detail in the winning box
+    # to distinguish the label from unmarked corners. Metadata confirms
+    # Doubao as the producer, not that the export retained a visible label.
+    _MIN_BRIGHT_NEUTRAL_FRACTION = 0.29
+    _MIN_MEAN_LAPLACIAN = 10.0
+    _BOX_PAD_LEFT_HEIGHT = 1.2
+    _BOX_PAD_RIGHT_HEIGHT = 0.7
+
+    @classmethod
+    def _outline_pixels_ok(cls, patch: NDArray[Any]) -> bool:
+        if patch.size == 0:
+            return False
+        patch = image_io.to_bgr(patch)
+        gray = cv2.cvtColor(patch, cv2.COLOR_BGR2GRAY)
+        saturation = cv2.cvtColor(patch, cv2.COLOR_BGR2HSV)[:, :, 1]
+        bright_neutral = float(np.mean((gray > 160) & (saturation < 60)))
+        edge_detail = float(np.mean(np.abs(cv2.Laplacian(gray, cv2.CV_32F))))
+        return bright_neutral >= cls._MIN_BRIGHT_NEUTRAL_FRACTION and edge_detail >= cls._MIN_MEAN_LAPLACIAN
+
+    def _scan(self, image: NDArray[Any] | None) -> TextMarkScan:
+        scan = super()._scan(image)
+        if image is None or scan.loc is None or scan.match_box is None or scan.score is None:
+            return scan
+        x0, y0, x1, y1 = scan.match_box
+        patch = image[
+            scan.loc.y + y0 : scan.loc.y + y1 + 1,
+            scan.loc.x + x0 : scan.loc.x + x1 + 1,
+        ]
+        if not self._outline_pixels_ok(patch):
+            return dataclasses.replace(scan, score=0.0)
+        return scan
+
+
 def _alpha_template() -> NDArray[Any] | None:
    """The bundled Doubao alpha template (float [0,1]), or None."""
    return _text_mark_engine.load_alpha_template(_CONFIG.asset_name)
@@ -132,11 +169,29 @@ def _template_match_score(box_mask: NDArray[Any], scale_base: int) -> float:
    return _text_mark_engine.template_match_score(box_mask, scale_base, _CONFIG)


-class DoubaoEngine(TextMarkEngine):
+class DoubaoEngine(TextMarkVariantsMixin):
    """Detect/localize the visible Doubao "豆包AI生成" watermark (locate -> mask; mask feeds the fill)."""

    def __init__(self) -> None:
        super().__init__(_CONFIG)
+        self._variants = (
+            _OutlinedDoubaoEngine(
+                dataclasses.replace(
+                    _CONFIG,
+                    asset_name="doubao_outline_alpha.png",
+                    width_frac=0.28,
+                    height_frac=0.13,
+                    scale_basis="width",
+                    alpha_width_frac=0.14,
+                    alpha_height_frac=0.14 * 0.75 * 77 / 346,
+                    max_saturation=255,
+                    detect_frontend="contrast",
+                    ladder=(0.95, 1.0, 1.05),
+                    detect_ncc_threshold=0.50,
+                    provenance_ncc_factor=0.25,
+                )
+            ),
+        )

    def _scan(self, image: NDArray[Any] | None) -> TextMarkScan:
        base = super()._scan(image)
@@ -179,12 +234,15 @@ class DoubaoEngine(TextMarkEngine):
        det = detection if detection is not None else self.detect(image)
        if not det.detected or det.match_box is None:
            return super().footprint_mask(image, force=False, dilate=dilate, detection=det)
-        alpha = _alpha_template()
+        engine = self.variant_engine(det)
+        if engine is not self:
+            return engine.footprint_mask(image, force=False, dilate=dilate, detection=det)
+        alpha = _text_mark_engine.load_alpha_template(engine.config.asset_name)
        if alpha is None:
            return super().footprint_mask(image, force=False, dilate=dilate, detection=det)

        radius = _FOOTPRINT_DILATE if dilate is None else max(0, dilate)
-        return self._aligned_alpha_mask(
+        return engine._aligned_alpha_mask(
            image,
            det,
            alpha,
@@ -957,7 +957,8 @@ def _integrity_clashes(
        clashes.append(
            "Conflicting AI-origin attributions from independent signals: "
            + " vs ".join(parts)
-            + " -- one provenance set was likely spoofed, transplanted, or laundered."
+            + " -- these fields may describe different stages of an edit chain; "
+            "the available evidence does not establish which attribution describes the final pixels."
        )

    # Rule 2: a camera-capture C2PA device next to an AI-generation marker. Only
@@ -967,13 +968,15 @@ def _integrity_clashes(
    # AI edit in ONE manifest, so the AI vendor is named only from that same
    # manifest (C2PA issuer + SynthID provenance) -- a legitimate edit chain, not a
    # spoof. An EXIF/XMP generator, IPTC field, TC260 AIGC label, or second
-    # manifest naming AI on a camera capture is the real laundering tell.
+    # manifest naming AI on a camera capture records a second stage whose relation
+    # to the capture cannot be resolved from these fields alone.
    independent_ai_marker = any(grp != _C2PA_MANIFEST_SOURCE for grp in source.values())
    if camera_label and camera_has_ai_marker and independent_ai_marker:
        vendors = ", ".join(sorted(set(ai_vendors.values()))) or "present"
        clashes.append(
            f"Camera-capture C2PA credentials ({camera_label}) coexist with AI-generation markers "
-            f"({vendors}) -- a genuine camera capture is not AI-generated, so the provenance is inconsistent."
+            f"({vendors}) -- these fields may describe a captured image followed by an AI edit; "
+            "the available evidence does not establish which stage describes the final pixels."
        )

    return clashes
@@ -172,7 +172,8 @@ class InvisibleEngine:
            cpu_offload: Offload model components to CPU between CUDA calls instead
                of keeping the whole pipeline in VRAM, at the cost of speed. Forces
                the face stack to offload instead of using automatic residency, on
-                every profile; reaches the global stack of qwen-zimage only.
+                every profile; reaches the global stack of qwen-zimage and
+                sdxl-zimage, not chroma-zimage.
                CUDA only.
        """

@@ -310,6 +311,15 @@ class InvisibleEngine:
        with Image.open(image_path) as opened:
            image = ImageOps.exif_transpose(opened)
        orig_size = image.size  # (width, height)
+        # Diffusion only processes RGB. Keep alpha at the original display geometry
+        # (after EXIF, before any resize) and rejoin it after all RGB postprocessing.
+        # Palette and colour-key PNGs can carry transparency without an A band.
+        alpha = (
+            image.convert("RGBA").getchannel("A")
+            if output_path.suffix.lower() in image_io.ALPHA_FORMATS
+            and ("A" in image.getbands() or "transparency" in image.info)
+            else None
+        )
        # Full-res original, kept for the adaptive-polish detail target (image is
        # reassigned to the resized copy below; PIL resize returns a new object).
        reference_pil = image
@@ -368,7 +378,7 @@ class InvisibleEngine:
            # Route those outputs through the same final resize so --no-adaptive-polish does
            # not silently change e.g. 1448x1086 into 1448x1080.
            needs_restore = target is not None or any(dimension % 8 for dimension in orig_size)
-            if humanize > 0.0 or unsharp > 0.0 or adaptive_polish or needs_restore:
+            if humanize > 0.0 or unsharp > 0.0 or adaptive_polish or needs_restore or alpha is not None:
                import cv2
                import numpy as np

@@ -389,7 +399,13 @@ class InvisibleEngine:
                    progress=self._progress_callback,
                )

-                if not image_io.imwrite(out_path, out_cv, display_tags_from=image_path, orientation_applied=True):
+                if not image_io.write_bgr_with_alpha(
+                    out_path,
+                    out_cv,
+                    np.asarray(alpha) if alpha is not None else None,
+                    display_tags_from=image_path,
+                    orientation_applied=True,
+                ):
                    raise OSError(f"failed to write postprocessed output: {out_path}")

            return out_path
@@ -56,6 +56,7 @@ from dataclasses import replace
 from typing import TYPE_CHECKING, Any

 import cv2
+import numpy as np

 from remove_ai_watermarks import _text_mark_engine
 from remove_ai_watermarks._text_mark_engine import TextMarkConfig, TextMarkDetection, TextMarkEngine
@@ -136,8 +137,12 @@ _CONFIG = TextMarkConfig(
 _LATIN_CONFIG = replace(
    _CONFIG,
    asset_name="kling_latin_alpha.png",
-    alpha_width_frac=0.095,
-    alpha_height_frac=0.028,
+    width_frac=0.22,
+    height_frac=0.07,
+    margin_x_frac=0.002,
+    margin_bottom_frac=0.002,
+    alpha_width_frac=0.14,
+    alpha_height_frac=0.032,
    ladder=(0.9, 1.0, 1.1),
    detect_ncc_threshold=0.40,
 )
@@ -153,12 +158,37 @@ def _glyph_silhouette() -> NDArray[Any] | None:
    return _text_mark_engine.glyph_silhouette(_CONFIG.asset_name)


-class KlingEngine(TextMarkEngine):
+class _KlingTextEngine(TextMarkEngine):
+    """Require glyph structure beyond a horizontal or vertical background edge."""
+
+    def _ladder_best(
+        self, image: NDArray[Any], loc: _text_mark_engine.TextMarkLocation
+    ) -> tuple[float, tuple[int, int, int, int] | None]:
+        score, box = super()._ladder_best(image, loc)
+        response = self._detect_response(image, loc)
+        silhouette = self._glyph_silhouette()
+        if box is None or response is None or silhouette is None:
+            return score, box
+        x, y, x1, y1 = box
+        patch = response[y : y1 + 1, x : x1 + 1].astype(np.float32)
+        template = cv2.resize(silhouette, (patch.shape[1], patch.shape[0]), interpolation=cv2.INTER_AREA)
+
+        def glyph_detail(array: NDArray[Any]) -> NDArray[Any]:
+            array = array.astype(np.float32)
+            return array - array.mean(axis=0, keepdims=True) - array.mean(axis=1, keepdims=True) + array.mean()
+
+        patch, template = glyph_detail(patch), glyph_detail(template)
+        denominator = float(np.linalg.norm(patch) * np.linalg.norm(template))
+        detail_score = float((patch * template).sum()) / denominator if denominator > 1e-6 else 0.0
+        return max(0.0, min(score, detail_score)), box
+
+
+class KlingEngine(_KlingTextEngine):
    """Detect/localize Kling AI's CJK and Latin 3.0 marks (locate -> mask -> fill)."""

    def __init__(self) -> None:
        super().__init__(_CONFIG)
-        self._latin = TextMarkEngine(_LATIN_CONFIG)
+        self._latin = _KlingTextEngine(_LATIN_CONFIG)

    def detect(self, image: NDArray[Any], *, provenance: bool = False) -> TextMarkDetection:
        """Return the strongest CJK or Latin Kling wordmark verdict."""
@@ -195,14 +225,15 @@ class KlingEngine(TextMarkEngine):
            return super().footprint_mask(image, force=force, dilate=dilate, detection=detection)

        det = detection if detection is not None else self.detect(image)
-        legacy = super().footprint_mask(image, force=False, dilate=dilate, detection=det)
+        if not det.detected:
+            return None
+        engine = self._latin if det.template_asset == _LATIN_CONFIG.asset_name else self
+        legacy = TextMarkEngine.footprint_mask(engine, image, force=False, dilate=dilate, detection=det)
        radius = _FOOTPRINT_DILATE if dilate is None else max(0, dilate)
        result = legacy
-        for config in (_CONFIG, _LATIN_CONFIG):
-            alpha = _text_mark_engine.load_alpha_template(config.asset_name)
-            if alpha is None:
-                continue
-            core = self._aligned_alpha_mask(
+        alpha = _text_mark_engine.load_alpha_template(engine.config.asset_name)
+        if alpha is not None:
+            core = engine._aligned_alpha_mask(
                image,
                det,
                alpha,
@@ -170,7 +170,7 @@ class LibLibPillEngine:
    """Detect the compact top-left pill only when LiblibAI is corroborated."""

    def __init__(self) -> None:
-        self._pill = PillEngine()
+        self._pill = PillEngine(include_label=False)

    def _raw(self, image: NDArray[Any]) -> PillDetection:
        return self._pill.detect(image)
@@ -0,0 +1,297 @@
+"""Model Context Protocol server over stdio, a thin wrapper around this package's CLI.
+
+``remove-ai-watermarks mcp`` speaks newline-delimited JSON-RPC 2.0 on stdin and
+stdout with the initialize handshake. Each tool call runs one CLI command in a
+child process of the same interpreter, so a tool returns exactly what that command
+would: its exit code, its output file and its text. Nothing but protocol messages
+reaches stdout, and a crash inside a command cannot corrupt the stream.
+
+The server needs no dependency beyond the standard library, so the bare package
+runs it through ``uvx remove-ai-watermarks mcp``. Commands that need an optional
+extra fail with the CLI's own install hint, returned as a tool error.
+"""
+
+from __future__ import annotations
+
+import contextlib
+import json
+import logging
+import os
+import subprocess
+import sys
+from collections.abc import Callable
+from pathlib import Path
+from typing import IO, Any, cast
+
+from remove_ai_watermarks import __version__
+from remove_ai_watermarks.video import VIDEO_EXTENSIONS
+
+log = logging.getLogger(__name__)
+
+# Newest first. A client asking for an unlisted version gets the newest, and
+# decides for itself whether it can continue.
+PROTOCOL_VERSIONS = ("2025-11-25", "2025-06-18", "2025-03-26", "2024-11-05")
+
+INSTRUCTIONS = (
+    "Identify and remove AI provenance marks from images and video the user generated "
+    "or edited themselves: C2PA, IPTC, EXIF and TC260 metadata, visible AI-generator "
+    "labels, and supported invisible watermarks. Do not use it on stock, agency, "
+    "marketplace or other third-party marks that protect someone else's asset. A "
+    "result of 'unknown' or exit code 2 means no supported signal was found, not "
+    "that the file is clean. Paths are read on the machine running the server."
+)
+
+_MODES = ("metadata", "visible", "all")
+
+TOOLS: list[dict[str, Any]] = [
+    {
+        "name": "identify",
+        "title": "Identify AI provenance",
+        "description": (
+            "Report which AI provenance signals an image or video carries: metadata "
+            "(C2PA, IPTC, EXIF, TC260), visible AI-generator labels, and supported local "
+            "invisible-watermark decoders. Read-only. Video is detected by file suffix."
+        ),
+        "inputSchema": {
+            "type": "object",
+            "properties": {
+                "path": {"type": "string", "description": "Image or video file to inspect."},
+                "check_visible": {
+                    "type": "boolean",
+                    "default": True,
+                    "description": "Run the pixel and visible-mark detectors. False reads metadata only.",
+                },
+            },
+            "required": ["path"],
+            "additionalProperties": False,
+        },
+        "annotations": {"readOnlyHint": True, "openWorldHint": False},
+    },
+    {
+        "name": "remove",
+        "title": "Remove AI watermarks from one file",
+        "description": (
+            "Write a cleaned copy of an image or video the user generated or edited. "
+            "mode=metadata strips AI metadata without touching pixels or re-encoding video; "
+            "mode=visible also removes a known visible AI label; for an image, mode=all adds "
+            "the invisible-watermark scrub where the installed extras and hardware allow it. "
+            "The source is never overwritten; the default output is <name>_clean beside it."
+        ),
+        "inputSchema": {
+            "type": "object",
+            "properties": {
+                "path": {"type": "string", "description": "Image or video file to clean."},
+                "mode": {"type": "string", "enum": list(_MODES), "default": "metadata"},
+                "output": {"type": "string", "description": "Output file. Must differ from the source."},
+            },
+            "required": ["path"],
+            "additionalProperties": False,
+        },
+        "annotations": {"readOnlyHint": False, "destructiveHint": False, "openWorldHint": False},
+    },
+    {
+        "name": "batch",
+        "title": "Remove AI watermarks from a directory",
+        "description": (
+            "Clean every supported image (media=image) or video (media=video) in a directory "
+            "the user owns, writing results to a separate directory (default <dir>_clean). "
+            "Modes as for the remove tool."
+        ),
+        "inputSchema": {
+            "type": "object",
+            "properties": {
+                "directory": {"type": "string", "description": "Directory of input files."},
+                "mode": {"type": "string", "enum": list(_MODES), "default": "metadata"},
+                "media": {"type": "string", "enum": ["image", "video"], "default": "image"},
+                "output_dir": {"type": "string", "description": "Output directory. Must differ from the input."},
+            },
+            "required": ["directory"],
+            "additionalProperties": False,
+        },
+        "annotations": {"readOnlyHint": False, "destructiveHint": False, "openWorldHint": False},
+    },
+]
+
+_TOOL_SCHEMAS = {tool["name"]: tool["inputSchema"] for tool in TOOLS}
+
+Runner = Callable[[list[str]], subprocess.CompletedProcess[str]]
+
+
+class ToolInputError(ValueError):
+    """A tool call whose arguments the server refuses before running anything."""
+
+
+def run_cli(args: list[str]) -> subprocess.CompletedProcess[str]:
+    """Run one CLI command in a child process of this interpreter."""
+    env = {**os.environ, "COLUMNS": "120", "NO_COLOR": "1", "TERM": "dumb"}
+    command = [sys.executable, "-m", "remove_ai_watermarks.cli", *args]
+    log.info("Running CLI command: %s", command)
+    completed = subprocess.run(  # noqa: S603 -- fixed interpreter, argument list, no shell
+        command,
+        stdin=subprocess.DEVNULL,
+        capture_output=True,
+        text=True,
+        env=env,
+        check=False,
+    )
+    log.info("CLI command exited %d: %s", completed.returncode, command)
+    return completed
+
+
+def _validated(tool: str, arguments: dict[str, Any]) -> dict[str, Any]:
+    """Check arguments against the tool's own input schema and fill its defaults."""
+    schema = _TOOL_SCHEMAS[tool]
+    properties: dict[str, dict[str, Any]] = schema["properties"]
+    unknown = sorted(set(arguments) - set(properties))
+    if unknown:
+        raise ToolInputError(f"unknown argument(s): {', '.join(unknown)}")
+    values: dict[str, Any] = {}
+    for key, spec in properties.items():
+        if key not in arguments:
+            if key in schema["required"]:
+                raise ToolInputError(f"'{key}' is required")
+            if "default" in spec:
+                values[key] = spec["default"]
+            continue
+        value = arguments[key]
+        if spec["type"] == "boolean" and not isinstance(value, bool):
+            raise ToolInputError(f"'{key}' must be a boolean")
+        if spec["type"] == "string" and (not isinstance(value, str) or not value.strip()):
+            raise ToolInputError(f"'{key}' must be a non-empty string")
+        if "enum" in spec and value not in spec["enum"]:
+            raise ToolInputError(f"'{key}' must be one of {', '.join(spec['enum'])}")
+        values[key] = value
+    return values
+
+
+def _is_video(path: str) -> bool:
+    return Path(path).suffix.lower() in VIDEO_EXTENSIONS
+
+
+def _identify_command(values: dict[str, Any]) -> tuple[list[str], str | None]:
+    path = values["path"]
+    prefix = ["video", "identify"] if _is_video(path) else ["identify"]
+    return [*prefix, "--json", *([] if values["check_visible"] else ["--no-visible"]), "--", path], None
+
+
+def _remove_command(values: dict[str, Any]) -> tuple[list[str], str | None]:
+    path, mode = values["path"], values["mode"]
+    source = Path(path)
+    output = values.get("output") or str(source.with_stem(source.stem + "_clean"))
+    if Path(output).resolve() == source.resolve():
+        raise ToolInputError("'output' must differ from the source; the server never overwrites an input")
+    command = ["metadata", "--remove"] if mode == "metadata" else [mode]
+    return [*(["video"] if _is_video(path) else []), *command, "-o", output, "--", path], output
+
+
+def _batch_command(values: dict[str, Any]) -> tuple[list[str], str | None]:
+    directory = values["directory"]
+    resolved = Path(directory).resolve()
+    output_dir = values.get("output_dir") or str(resolved.with_name(resolved.name + "_clean"))
+    if Path(output_dir).resolve() == resolved:
+        raise ToolInputError("'output_dir' must differ from the input directory")
+    command = ["video", "batch"] if values["media"] == "video" else ["batch"]
+    return [*command, "--mode", values["mode"], "-o", output_dir, "--", directory], output_dir
+
+
+_COMMANDS: dict[str, Callable[[dict[str, Any]], tuple[list[str], str | None]]] = {
+    "identify": _identify_command,
+    "remove": _remove_command,
+    "batch": _batch_command,
+}
+
+
+def call_tool(name: str, arguments: dict[str, Any], runner: Runner = run_cli) -> dict[str, Any]:
+    """Run one tool and return an MCP ``tools/call`` result."""
+    try:
+        args, output = _COMMANDS[name](_validated(name, arguments))
+    except ToolInputError as exc:
+        return {"content": [{"type": "text", "text": f"Invalid arguments: {exc}"}], "isError": True}
+    completed = runner(args)
+    code = completed.returncode
+    text = completed.stdout.strip()
+    stderr = completed.stderr.strip()[-4000:]
+    structured: dict[str, Any] = {"command": ["remove-ai-watermarks", *args], "exit_code": code, "output": output}
+    if "--json" in args and code == 0:
+        with contextlib.suppress(json.JSONDecodeError):
+            structured["report"] = json.loads(text)
+    if "report" not in structured:
+        structured["stdout"] = text
+    if stderr:
+        structured["stderr"] = stderr
+    # The CLI's own exit 2 means no supported signal, a result. Click also exits 2 on
+    # a usage error, such as a missing file, and prints its usage line on stderr.
+    no_signal = code == 2 and "Usage:" not in stderr
+    is_error = code != 0 and not no_signal
+    summary = f"exit code {code}"
+    if no_signal:
+        summary += " (no supported signal found; this is unknown, not proof the file is clean)"
+    body = "\n\n".join(part for part in (summary, text, stderr if is_error else "") if part)
+    return {"content": [{"type": "text", "text": body}], "structuredContent": structured, "isError": is_error}
+
+
+def _error(request_id: object, code: int, message: str) -> dict[str, Any]:
+    return {"jsonrpc": "2.0", "id": request_id, "error": {"code": code, "message": message}}
+
+
+def _object(value: object) -> dict[str, Any] | None:
+    """A decoded JSON object, or ``None`` for any other JSON value."""
+    return cast("dict[str, Any]", value) if isinstance(value, dict) else None
+
+
+def handle(raw: object, runner: Runner = run_cli) -> dict[str, Any] | None:
+    """Answer one JSON-RPC message; ``None`` for notifications and client responses."""
+    message = _object(raw)
+    if message is None or message.get("jsonrpc") != "2.0":
+        return _error(None, -32600, "Invalid Request")
+    method = message.get("method")
+    if "id" not in message:
+        return None  # a notification or a response to nothing this server sent
+    request_id: object = message["id"]
+    if not isinstance(method, str):
+        return _error(request_id, -32600, "Invalid Request")
+    params = _object(message.get("params") or {})
+    if params is None:
+        return _error(request_id, -32602, "params must be an object")
+
+    if method == "initialize":
+        requested = params.get("protocolVersion")
+        result: dict[str, Any] = {
+            "protocolVersion": requested if requested in PROTOCOL_VERSIONS else PROTOCOL_VERSIONS[0],
+            "capabilities": {"tools": {"listChanged": False}},
+            "serverInfo": {"name": "remove-ai-watermarks", "title": "Remove AI Watermarks", "version": __version__},
+            "instructions": INSTRUCTIONS,
+        }
+    elif method == "ping":
+        result = {}
+    elif method == "tools/list":
+        result = {"tools": TOOLS}
+    elif method == "tools/call":
+        name = params.get("name")
+        arguments = _object(params.get("arguments") or {})
+        if not isinstance(name, str) or name not in _COMMANDS:
+            return _error(request_id, -32602, f"Unknown tool: {name}")
+        if arguments is None:
+            return _error(request_id, -32602, "arguments must be an object")
+        result = call_tool(name, arguments, runner)
+    else:
+        return _error(request_id, -32601, f"Method not found: {method}")
+    return {"jsonrpc": "2.0", "id": request_id, "result": result}
+
+
+def serve(stdin: IO[str] | None = None, stdout: IO[str] | None = None, runner: Runner = run_cli) -> None:
+    """Serve newline-delimited JSON-RPC until stdin closes."""
+    source: IO[str] = sys.stdin if stdin is None else stdin
+    sink: IO[str] = sys.stdout if stdout is None else stdout
+    for line in source:
+        if not line.strip():
+            continue
+        try:
+            message = json.loads(line)
+        except json.JSONDecodeError:
+            response: dict[str, Any] | None = _error(None, -32700, "Parse error")
+        else:
+            response = handle(message, runner)
+        if response is not None:
+            sink.write(json.dumps(response, ensure_ascii=False) + "\n")
+            sink.flush()
@@ -85,6 +85,11 @@ AI_KEYWORDS: tuple[str, ...] = (
 # Reference: https://spec.c2pa.org/specifications/specifications/2.1/specs/C2PA_Specification.html
 C2PA_UUID: bytes = bytes.fromhex("d8fec3d61b0e483c92975828877ec481")

+# Kling's encoder writes an H.264 ``user_data_unregistered`` SEI message (payload
+# type 5) whose payload is this UUID followed by ``kling-ai``. It lives in the coded
+# stream inside ``mdat``, where no metadata box walk reaches it.
+KLING_SEI_TAG: bytes = bytes.fromhex("91ca60614aee385486142d5f73f4ae2e") + b"kling-ai"
+

 def c2pa_marker_in(data: bytes) -> bool:
    """True if ``data`` carries a real C2PA manifest marker, not just an
@@ -1430,6 +1435,10 @@ def get_ai_metadata(image_path: Path) -> dict[str, str]:
    if app_generator:
        result.setdefault("app_aigc", f"App AIGC disclosure ({app_generator})")

+    # Kling's encoder tag, an H.264 SEI message in the first video sample.
+    if KLING_SEI_TAG in app_scan:
+        result.setdefault("kling_sei", "Kling AI encoder tag (H.264 SEI kling-ai)")
+
    # xAI / Grok signature pair in EXIF or an equivalent container-native field.
    if xai_signature(image_path):
        result.setdefault("xai_signature", "xAI/Grok signature (UUID + Signature blob)")
@@ -38,10 +38,12 @@ status was not adjudicated, and 1200 non-overlapping no-signal controls:

 from __future__ import annotations

+from dataclasses import replace
 from typing import TYPE_CHECKING, Any

 from remove_ai_watermarks import _text_mark_engine
-from remove_ai_watermarks._text_mark_engine import TextMarkConfig, TextMarkEngine
+from remove_ai_watermarks._text_mark_engine import TextMarkConfig, TextMarkDetection
+from remove_ai_watermarks._text_mark_variants import GlyphTextMarkEngine, TextMarkVariantsMixin

 if TYPE_CHECKING:
    from numpy.typing import NDArray
@@ -101,8 +103,62 @@ def _alpha_template() -> NDArray[Any] | None:
    return _text_mark_engine.load_alpha_template(_CONFIG.asset_name)


-class MicrosoftEngine(TextMarkEngine):
+class MicrosoftEngine(TextMarkVariantsMixin):
    """Detect/localize the measured Microsoft top-right AI badge."""

    def __init__(self) -> None:
        super().__init__(_CONFIG)
+        self._variants = (
+            GlyphTextMarkEngine(
+                replace(
+                    _CONFIG,
+                    asset_name="microsoft_text_alpha.png",
+                    width_frac=0.20,
+                    height_frac=0.065,
+                    margin_x_frac=0.002,
+                    margin_bottom_frac=0.002,
+                    scale_basis="width",
+                    alpha_width_frac=0.14,
+                    alpha_height_frac=0.14 * 49 / 339,
+                    detect_frontend="contrast",
+                    ladder=(0.95, 1.0, 1.05),
+                    detect_ncc_threshold=0.60,
+                    provenance_ncc_factor=1.0,
+                    max_saturation=255,
+                )
+            ),
+        )
+
+    def footprint_mask(
+        self,
+        image: NDArray[Any] | None,
+        *,
+        force: bool = False,
+        dilate: int | None = None,
+        detection: TextMarkDetection | None = None,
+    ) -> NDArray[Any] | None:
+        """Cover the enclosing beige pill for the text-only layout."""
+        if force or image is None or image.size == 0:
+            return super().footprint_mask(image, force=force, dilate=dilate, detection=detection)
+        det = detection if detection is not None else self.detect(image)
+        engine = self.variant_engine(det)
+        if engine is self or not det.detected or det.match_box is None:
+            return super().footprint_mask(image, force=False, dilate=dilate, detection=det)
+
+        from remove_ai_watermarks import region_eraser
+
+        loc = engine.locate(image)
+        x0, y0, x1, y1 = det.match_box
+        pad_x = int(image.shape[1] * 0.012)
+        pad_y = int(image.shape[1] * 0.014)
+        box = (
+            max(0, loc.x + x0 - pad_x),
+            max(0, loc.y + y0 - pad_y),
+            x1 - x0 + 1 + 2 * pad_x,
+            y1 - y0 + 1 + 2 * pad_y,
+        )
+        return region_eraser.boxes_to_mask(
+            image.shape[:2],
+            [box],
+            dilate=0 if dilate is None else max(0, dilate),
+        )
@@ -3,8 +3,7 @@
 OpenArt (openart.ai) stamps free-plan generations with a semi-transparent white
 "OpenArt" wordmark (a bowtie/infinity icon followed by the brand name) placed
 **near the center of the frame**, not in a corner like every other text mark
-this repository detects. Confirmed from one customer-reported production case
-(raiw-app, 2026-09-11, 1528x2712 portrait output): the mark sat roughly centered
+this repository detects. In the observed portrait layout, the mark sat roughly centered
 over the subject's torso, and neither C2PA, EXIF AI tags, nor the China AIGC/TC260
 label were present -- OpenArt apparently ships no machine-readable provenance
 signal alongside this visible mark, so a bare export was previously reported as
@@ -1,18 +1,20 @@
-"""Jimeng-basic 'AI生成' pill: a CAPTURE-LESS visible mark (issue #54).
+"""Jimeng-basic ``AI生成`` pill visible mark (issue #54).

 The Jimeng free-tier TC260 label is a rounded pill with 'AI生成' in the TOP-LEFT
 corner -- distinct from the ``jimeng`` "★ 即梦AI" wordmark (bottom-right). It has no
-captured alpha map, so unlike the other marks it is detected purely by a synthetic
-silhouette; like every mark it is then removed by the shared localize -> fill:
+captured alpha map for its outline. Detection therefore has two paths before the
+shared localize -> fill step:

-  * Detect: edge-NCC of a font-rendered SILHOUETTE (``assets/jimeng_pill.png``,
+  * A high-confidence contrast template for the measured solid ``AI生成`` label
+    (``assets/jimeng_label_alpha.png``). With Jimeng provenance this verified
+    label can bypass the flat-background safety gate.
+  * Edge-NCC of a font-rendered SILHOUETTE (``assets/jimeng_pill.png``,
    synthetic, data-safe -- see ``scripts/render_pill_silhouette.py``) against the
-    top-left ROI, at the pill's known width fraction. The calibrated
-    ``_DETECT_THRESHOLD`` is 0.22.
-  * Remove: place the pill footprint at the matched location and inpaint it
-    (MI-GAN / cv2 via the registry). Quality comes from the inpaint backend, so the
-    silhouette need not be pixel-accurate -- which is why a synthetic render is
-    sufficient and no source-derived asset is committed.
+    top-left ROI, at the pill's known width fraction. Its calibrated
+    ``_DETECT_THRESHOLD`` is 0.22 and it retains the provenance and flat-background
+    product gates.
+  * Remove: mask the verified label match or the stable outline footprint and
+    inpaint it (MI-GAN / cv2 via the registry).

 Geometry uses width ~0.161*W,
 height ~0.091*W, top-left, margins ~0.02-0.05.
@@ -27,6 +29,7 @@ import cv2
 import numpy as np

 from remove_ai_watermarks import image_io
+from remove_ai_watermarks._text_mark_engine import TextMarkConfig, TextMarkDetection, TextMarkEngine

 if TYPE_CHECKING:
    from numpy.typing import NDArray
@@ -74,6 +77,32 @@ class PillDetection(NamedTuple):
    detected: bool
    confidence: float
    region: tuple[int, int, int, int]  # x, y, w, h of the matched pill
+    verified_label: bool = False
+
+
+_LABEL_CONFIG = TextMarkConfig(
+    name="Jimeng AI生成 pill label",
+    asset_name="jimeng_label_alpha.png",
+    corner="tl",
+    margin_floor=2,
+    width_frac=0.34,
+    height_frac=0.14,
+    margin_x_frac=0.0,
+    margin_bottom_frac=0.0,
+    max_saturation=255,
+    logo_min_luma=100,
+    tophat_delta=8,
+    morph_open_size=3,
+    detect_min_coverage=0.01,
+    detect_ncc_threshold=0.55,
+    detect_frontend="contrast",
+    scale_basis="width",
+    ladder=(0.95, 1.0, 1.05),
+    alpha_width_frac=0.14,
+    alpha_height_frac=0.14 * 128 / 353,
+    min_gw=24,
+    provenance_ncc_factor=1.0,
+)


 def _load_silhouette() -> NDArray[Any] | None:
@@ -92,7 +121,18 @@ def _grad(gray: NDArray[Any]) -> NDArray[Any]:


 class PillEngine:
-    """Detect + build the removal mask for the top-left 'AI生成' pill (edge-NCC of a synthetic silhouette)."""
+    """Detect the top-left ``AI生成`` label or its synthetic pill outline."""
+
+    def __init__(self, *, include_label: bool = True) -> None:
+        self._label = TextMarkEngine(_LABEL_CONFIG) if include_label else None
+
+    @staticmethod
+    def _label_box(detection: TextMarkDetection) -> tuple[int, int, int, int] | None:
+        if not detection.detected or detection.match_box is None:
+            return None
+        x, y, _width, _height = detection.region
+        x0, y0, x1, y1 = detection.match_box
+        return x + x0, y + y0, x1 - x0 + 1, y1 - y0 + 1

    def _match(self, image: NDArray[Any]) -> tuple[float, tuple[int, int, int, int]] | None:
        sil = _load_silhouette()
@@ -113,12 +153,33 @@ class PillEngine:
        _, score, _, loc = cv2.minMaxLoc(res)
        return float(score), (int(loc[0]), int(loc[1]), tw, th)

-    def detect(self, image: NDArray[Any]) -> PillDetection:
-        m = self._match(image)
-        if m is None:
-            return PillDetection(False, 0.0, (0, 0, 0, 0))
-        score, box = m
-        return PillDetection(score >= _DETECT_THRESHOLD, score, box)
+    def _detection(
+        self,
+        match: tuple[float, tuple[int, int, int, int]] | None,
+        label: TextMarkDetection,
+    ) -> PillDetection:
+        label_box = self._label_box(label)
+        if match is None:
+            return PillDetection(label.detected, label.confidence, label_box or label.region, label.detected)
+        score, box = match
+        return PillDetection(
+            score >= _DETECT_THRESHOLD or label.detected,
+            max(score, label.confidence),
+            label_box if label.detected and label_box is not None else box,
+            label.detected,
+        )
+
+    def detect(self, image: NDArray[Any], *, provenance: bool = False) -> PillDetection:
+        """Detect the outline, plus the solid label when Jimeng is corroborated."""
+        label_scan = self._label.detect(image) if self._label is not None else TextMarkDetection()
+        label = label_scan if provenance else TextMarkDetection()
+        return self._detection(self._match(image), label)
+
+    def detect_both(self, image: NDArray[Any]) -> tuple[PillDetection, PillDetection]:
+        """Return strict outline and Jimeng-corroborated label verdicts from one scan."""
+        match = self._match(image)
+        label = self._label.detect(image) if self._label is not None else TextMarkDetection()
+        return self._detection(match, TextMarkDetection()), self._detection(match, label)

    def _footprint_box(self, image: NDArray[Any]) -> tuple[int, int, int, int] | None:
        h, w = image.shape[:2]
@@ -152,19 +213,34 @@ class PillEngine:
        """True when the top-left footprint is flat enough for an invisible inpaint."""
        return self.footprint_texture(image) <= thresh

-    def footprint_mask(self, image: NDArray[Any], *, force: bool = False) -> NDArray[Any] | None:
+    def footprint_mask(
+        self,
+        image: NDArray[Any],
+        *,
+        force: bool = False,
+        detection: PillDetection | None = None,
+    ) -> NDArray[Any] | None:
        """Full-frame uint8 mask (255 = pill) over the pill's known top-left region.

-        Uses stable GEOMETRY (a generous fixed box), not the NCC match position: the
-        synthetic silhouette localizes only approximately, so a match-positioned mask
-        leaves outline residue, while the top-left corner is negative space, so a
-        generous geometric box removes the pill cleanly and harmlessly. The caller
+        A verified label uses its detector-aligned box with measured padding. The
+        weaker synthetic outline uses stable geometry because its NCC position is
+        only approximate and a match-positioned mask leaves rim residue. The caller
        gates on :meth:`detect`, so a clean corner is never masked. ``force`` is
-        accepted for a uniform engine signature but ignored (the geometry box is
-        fixed regardless)."""
+        accepted for a uniform engine signature but ignored."""
        if image is None or image.size == 0:
            return None
-        box = self._footprint_box(image)
+        if detection is not None and detection.verified_label:
+            x, y, width, height = detection.region
+            pad_x = int(image.shape[1] * 0.025)
+            pad_y = int(image.shape[1] * 0.014)
+            box = (
+                max(0, x - pad_x),
+                max(0, y - pad_y),
+                min(image.shape[1], x + width + pad_x),
+                min(image.shape[0], y + height + pad_y),
+            )
+        else:
+            box = self._footprint_box(image)
        if box is None:
            return None
        # Same primitive the shared fill uses, rather than a private zeros/fill copy.
@@ -25,10 +25,12 @@ signal via the registry.

 from __future__ import annotations

+from dataclasses import replace
 from typing import TYPE_CHECKING, Any

 from remove_ai_watermarks import _text_mark_engine, image_io
-from remove_ai_watermarks._text_mark_engine import TextMarkConfig, TextMarkDetection, TextMarkEngine
+from remove_ai_watermarks._text_mark_engine import TextMarkConfig, TextMarkDetection
+from remove_ai_watermarks._text_mark_variants import GlyphTextMarkEngine, TextMarkVariantsMixin

 if TYPE_CHECKING:
    from numpy.typing import NDArray
@@ -107,11 +109,27 @@ def _template_match_score(box_mask: NDArray[Any], scale_base: int) -> float:
    return _text_mark_engine.template_match_score(box_mask, scale_base, _CONFIG)


-class SamsungEngine(TextMarkEngine):
+class SamsungEngine(TextMarkVariantsMixin):
    """Detect/localize the visible Samsung Galaxy AI text mark (locate -> mask; mask feeds the fill)."""

    def __init__(self) -> None:
        super().__init__(_CONFIG)
+        self._variants = (
+            GlyphTextMarkEngine(
+                replace(
+                    _CONFIG,
+                    asset_name="samsung_ko_alpha.png",
+                    width_frac=0.28,
+                    height_frac=0.065,
+                    alpha_width_frac=0.20,
+                    alpha_height_frac=0.20 * 49 / 344,
+                    ladder=(0.95, 1.0, 1.05),
+                    max_saturation=255,
+                    detect_frontend="contrast",
+                    detect_ncc_threshold=0.46,
+                )
+            ),
+        )

    def footprint_mask(
        self,
@@ -134,10 +152,29 @@ class SamsungEngine(TextMarkEngine):

        image = image_io.to_bgr(image)
        det = detection if detection is not None else self.detect(image)
-        alpha = _alpha_template()
+        engine = self.variant_engine(det)
+        radius = _FOOTPRINT_DILATE if dilate is None else max(0, dilate)
+        if engine is not self and det.detected and det.match_box is not None:
+            from remove_ai_watermarks import region_eraser
+
+            x0, y0, x1, y1 = det.match_box
+            loc = engine.locate(image)
+            glyph_height = y1 - y0 + 1
+            return region_eraser.boxes_to_mask(
+                image.shape[:2],
+                [
+                    (
+                        max(0, loc.x + x0 - round(1.3 * glyph_height)),
+                        loc.y + y0,
+                        x1 - x0 + 1 + round(1.3 * glyph_height),
+                        glyph_height,
+                    )
+                ],
+                dilate=radius,
+            )
+        alpha = _text_mark_engine.load_alpha_template(engine.config.asset_name)
        if alpha is not None:
-            radius = _FOOTPRINT_DILATE if dilate is None else max(0, dilate)
-            aligned = self._aligned_alpha_mask(
+            aligned = engine._aligned_alpha_mask(
                image,
                det,
                alpha,
@@ -377,6 +377,8 @@ def _platform_from_video_metadata(markers: dict[str, str]) -> str | None:
        return _tc260_video_platform(markers.get("aigc_producer", "")) or (
            "China AIGC-labeled content (TC260 standard)"
        )
+    if "kling_sei" in markers:
+        return _tc260_video_platform("kling")
    return None


@@ -135,7 +135,7 @@ def load_video_vae_runtime(
    dtype = torch.float16 if resolved_device == "cuda" else torch.float32
    log.info("Loading %s on %s", model, resolved_device)
    vae = AutoencoderKL.from_pretrained(model, torch_dtype=dtype).to(resolved_device)
-    vae.eval()
+    vae.train(False)
    vae.enable_slicing()
    scaling_factor = float(vae.config.scaling_factor)
    log.info("Latent scaling factor %.5f", scaling_factor)
@@ -24,13 +24,15 @@ Entries:
  - ``qwen`` -- Alibaba Cloud Qwen "千问AI生成" text or three-lobe symbol, bottom-right.
  - ``kling`` -- Kuaishou Kling AI "可灵AI 3.0" / "KlingAI 3.0" strip, bottom-right.
  - ``yuanbao`` -- Tencent Yuanbao "元宝 / AI生成" two-line mark, bottom-right.
-  - ``samsung`` -- Samsung Galaxy AI "Contenuti generati dall'AI" strip, bottom-left.
-  - ``jimeng_pill`` -- Jimeng-basic "AI生成" pill, top-left (capture-less).
+  - ``dola`` -- ByteDance Dola "Dola AI" image disclosure, bottom-right.
+  - ``workbuddy`` -- Tencent WorkBuddy "AI生成 / WORKBUDDY" disclosure, bottom-right.
+  - ``samsung`` -- Samsung Galaxy AI Italian or Korean disclosure, bottom-left.
+  - ``jimeng_pill`` -- Jimeng-basic "AI生成" pill or solid label, top-left.
  - ``runninghub`` -- RunningHub "RunningHub AI生成" text, top-left (gray front-end).
  - ``baidu`` -- Baidu "百度 AI生成" text + white tag, bottom-right.
  - ``liblib`` -- LiblibAI "LiblibAI" wordmark, bottom-center.
  - ``liblib_pill`` -- LiblibAI compact "AI生成" pill, top-left.
-  - ``microsoft`` -- one measured Microsoft white AI-badge variant, top-right.
+  - ``microsoft`` -- measured Microsoft white AI badge or beige text pill, top-right.
  - ``generic_ai_label`` -- brand-less bare "AI生成" TC260 label (no vendor wordmark),
    bottom-right; the fallback for a compliance stamp not tied to any tuned mark above.
  - ``openart`` -- OpenArt "OpenArt" wordmark (bowtie/infinity icon + brand name),
@@ -448,6 +450,8 @@ _ENGINE_CLASS: dict[str, tuple[str, str]] = {
    "jimeng": ("jimeng_engine", "JimengEngine"),
    "qwen": ("qwen_engine", "QwenEngine"),
    "kling": ("kling_engine", "KlingEngine"),
+    "dola": ("dola_engine", "DolaEngine"),
+    "workbuddy": ("workbuddy_engine", "WorkBuddyEngine"),
    "yuanbao": ("yuanbao_engine", "YuanbaoEngine"),
    "samsung": ("samsung_engine", "SamsungEngine"),
    "jimeng_pill": ("pill_engine", "PillEngine"),
@@ -641,29 +645,44 @@ def _text_mark(
    )


-# ── Capture-less mark: the Jimeng-basic "AI生成" pill (top-left) ──
-# Detection is edge-NCC of a synthetic silhouette; the mask is a fixed top-left
-# geometry box (see pill_engine). Removal is the same localize -> fill as the rest.
+# ── Jimeng-basic "AI生成" pill or solid label (top-left) ──
+# The weak outline uses synthetic edge-NCC and fixed geometry. The verified solid
+# label uses its own contrast template and detector-aligned mask (see pill_engine).
 def _pill_detect(image: NDArray[Any], *, provenance: bool = False) -> MarkDetection:
-    del provenance  # the pill detector is provenance-independent; its relaxation lives entirely in _keep_pill
-    d = _engine("jimeng_pill").detect(image)
-    return MarkDetection("jimeng_pill", "Jimeng AI生成 pill", "top-left", d.detected, d.confidence, d.region)
+    d = _engine("jimeng_pill").detect(image, provenance=provenance)
+    return MarkDetection(
+        "jimeng_pill",
+        "Jimeng AI生成 pill",
+        "top-left",
+        d.detected,
+        d.confidence,
+        d.region,
+        engine_detection=d,
+    )


 def _pill_detect_both(image: NDArray[Any]) -> tuple[MarkDetection, MarkDetection]:
-    # The pill detector is provenance-independent (`_pill_detect` discards the flag), so
-    # one call answers both levels. MarkDetection is frozen, so sharing it is safe.
-    d = _pill_detect(image)
-    return d, d
+    strict, relaxed = _engine("jimeng_pill").detect_both(image)
+
+    def wrap(detection: Any) -> MarkDetection:
+        return MarkDetection(
+            "jimeng_pill",
+            "Jimeng AI生成 pill",
+            "top-left",
+            detection.detected,
+            detection.confidence,
+            detection.region,
+            engine_detection=detection,
+        )
+
+    return wrap(strict), wrap(relaxed)


 def _pill_mask(
    image: NDArray[Any], *, force: bool = False, detection: MarkDetection | None = None
 ) -> NDArray[Any] | None:
-    # The pill mask is a fixed top-left geometry box, independent of the detection;
-    # accepted for the uniform _mask signature.
-    del detection
-    return _engine("jimeng_pill").footprint_mask(image, force=force)
+    engine_detection = detection.engine_detection if detection is not None else None
+    return _engine("jimeng_pill").footprint_mask(image, force=force, detection=engine_detection)


 def _pill_features(image: NDArray[Any]) -> dict[str, float]:
@@ -732,6 +751,24 @@ _REGISTRY: tuple[KnownMark, ...] = (
        platform="Tencent Yuanbao (visible 元宝 / AI生成 mark detected)",
        manufacturer="tencent",
    ),
+    _text_mark(
+        "dola",
+        "Dola AI image wordmark",
+        "bottom-right",
+        platform="Dola AI (visible Dola AI disclosure)",
+        manufacturer="bytedance",
+        label_regime=None,
+        provenance_signals=(),
+    ),
+    _text_mark(
+        "workbuddy",
+        "AI生成 / WORKBUDDY disclosure",
+        "bottom-right",
+        platform="WorkBuddy (visible AI content disclosure)",
+        manufacturer="tencent",
+        label_regime=None,
+        provenance_signals=(),
+    ),
    # Samsung Galaxy AI is a device editing marker (samsung_genai), not a TC260 label.
    _text_mark(
        "samsung",
@@ -940,8 +977,14 @@ def _pill_suppressors() -> set[str]:
    return {m.key for m in _REGISTRY if m.manufacturer == pill.manufacturer and m.product != pill.product}


-def _keep_pill(keys: set[str], *, provenance: frozenset[str], footprint_flat: bool) -> bool:
-    """Whether to auto-remove the capture-less 'AI生成' pill given the fired marks.
+def _keep_pill(
+    keys: set[str],
+    *,
+    provenance: frozenset[str],
+    footprint_flat: bool,
+    verified_label: bool = False,
+) -> bool:
+    """Whether to auto-remove the ``AI生成`` pill or verified label.

    Pure decision (the flatness feature is precomputed at perception time and passed
    in). The pill detector is weak and metadata/intent confirms the platform, not pill
@@ -950,23 +993,24 @@ def _keep_pill(keys: set[str], *, provenance: frozenset[str], footprint_flat: bo
      * bottom-right "★ 即梦AI" wordmark fired -> ~94% precise, and it survives
        metadata-STRIPPED images: remove the pill unrestricted;
      * TC260 metadata confirms Jimeng (``"jimeng" in provenance``, no wordmark) -> remove ONLY when the
-        top-left footprint is flat enough for an invisible fill (``footprint_flat``),
-        so real flat-scene pills (and harmless flat false fires) are cleaned while the
-        damaging textured false fires are left untouched.
+        solid label template verifies the text, or when the weaker outline's top-left
+        footprint is flat enough for an invisible fill (``footprint_flat``). Real
+        flat-scene pills are cleaned while damaging textured false fires are left.
    A Doubao image is TC260 too but is not Jimeng-basic, so its mark suppresses the
-    pill. Marks from other TC260 manufacturers are unrelated and do not participate in
-    this ByteDance-family decision.
+    pill. The independently attributed LiblibAI pill also wins over the generic
+    ``AI生成`` label shape. Marks from other TC260 manufacturers are unrelated and do
+    not participate in this ByteDance-family decision.
    No confirmation at all -> never remove (blocks false fires on non-Jimeng content).

    The suppressor set is derived from the manufacturer and product fields rather than
    from the shared TC260 standard. Marks from Google, Samsung, or another TC260
    manufacturer cannot enable or veto this ByteDance-specific arm."""
-    if _pill_suppressors() & keys:
+    if "liblib_pill" in keys or _pill_suppressors() & keys:
        return False
    if "jimeng" in keys:
        return True
    if "jimeng" in provenance:
-        return footprint_flat
+        return footprint_flat or verified_label
    return False


@@ -990,6 +1034,8 @@ def _build_candidates(image: NDArray[Any]) -> list[Candidate]:
            continue
        strict, relaxed = m.detect_both(image)
        feats = m.features(image) if (strict.detected or relaxed.detected) else {}
+        if relaxed.engine_detection is not None and getattr(relaxed.engine_detection, "verified_label", False):
+            feats["verified_label"] = 1.0
        cands.append(
            Candidate(
                m.key,
@@ -1010,7 +1056,7 @@ def decide(candidates: list[Candidate], context: Context) -> list[Decision]:

    All policy lives here, in one place: per-mark trust resolution (:func:`resolve_trust`,
    which needs the strict-detected siblings for ``auto`` cross-mark corroboration) and
-    the capture-less pill gate (:func:`_keep_pill`). No image, no I/O -- so it is unit-testable in isolation and
+    the Jimeng pill/label gate (:func:`_keep_pill`). No image, no I/O -- so it is unit-testable in isolation and
    the same decision drives every caller."""
    strict_keys = {c.key for c in candidates if c.detected_strict}
    fired: list[Decision] = []
@@ -1026,7 +1072,13 @@ def decide(candidates: list[Candidate], context: Context) -> list[Decision]:
    if "jimeng_pill" in keys:
        pill = next(d for d in fired if d.candidate.key == "jimeng_pill")
        flat = bool(pill.candidate.features.get("footprint_flat", 0.0))
-        if not _keep_pill(keys, provenance=context.provenance, footprint_flat=flat):
+        verified_label = bool(pill.candidate.features.get("verified_label", 0.0))
+        if not _keep_pill(
+            keys,
+            provenance=context.provenance,
+            footprint_flat=flat,
+            verified_label=verified_label,
+        ):
            fired = [d for d in fired if d.candidate.key != "jimeng_pill"]
    return fired

@@ -0,0 +1,35 @@
+"""Detect the corner AI生成 / WORKBUDDY disclosure without inferring a base model."""
+
+from remove_ai_watermarks._text_mark_engine import TextMarkConfig
+from remove_ai_watermarks._text_mark_variants import BottomRightBoxTextMarkEngine
+
+_CONFIG = TextMarkConfig(
+    name="WorkBuddy AI disclosure",
+    asset_name="workbuddy_alpha.png",
+    corner="br",
+    margin_floor=4,
+    width_frac=0.22,
+    height_frac=0.13,
+    margin_x_frac=0.002,
+    margin_bottom_frac=0.002,
+    max_saturation=255,
+    logo_min_luma=110,
+    tophat_delta=8,
+    morph_open_size=3,
+    detect_min_coverage=0.01,
+    detect_ncc_threshold=0.60,
+    alpha_width_frac=0.12,
+    alpha_height_frac=0.12 * 177 / 351,
+    detect_frontend="tophat",
+    scale_basis="width",
+    ladder=(0.95, 1.0, 1.05),
+    provenance_ncc_factor=1.0,
+    min_gw=24,
+)
+
+
+class WorkBuddyEngine(BottomRightBoxTextMarkEngine):
+    """Detect and localize the two-line WorkBuddy disclosure."""
+
+    def __init__(self) -> None:
+        super().__init__(_CONFIG)
@@ -22,6 +22,7 @@ covered by this silhouette.

 from __future__ import annotations

+from dataclasses import replace
 from typing import TYPE_CHECKING, Any

 from remove_ai_watermarks import _text_mark_engine
@@ -29,6 +30,7 @@ from remove_ai_watermarks._text_mark_engine import (
    BottomRightAnchoredEngine,
    TextMarkConfig,
 )
+from remove_ai_watermarks._text_mark_variants import BottomRightBoxTextMarkEngine, TextMarkVariantsMixin

 if TYPE_CHECKING:
    from numpy.typing import NDArray
@@ -79,7 +81,7 @@ def _alpha_template() -> NDArray[Any] | None:
    return _text_mark_engine.load_alpha_template(_CONFIG.asset_name)


-class YuanbaoEngine(BottomRightAnchoredEngine):
+class YuanbaoEngine(TextMarkVariantsMixin, BottomRightAnchoredEngine):
    """Detect and localize the bottom-right Yuanbao mark."""

    _ANCHOR_MAX_RIGHT = 0.04
@@ -87,3 +89,36 @@ class YuanbaoEngine(BottomRightAnchoredEngine):

    def __init__(self) -> None:
        super().__init__(_CONFIG)
+        self._variants = (
+            BottomRightBoxTextMarkEngine(
+                replace(
+                    _CONFIG,
+                    asset_name="yuanbao_compact_alpha.png",
+                    width_frac=0.28,
+                    height_frac=0.13,
+                    scale_basis="width",
+                    alpha_width_frac=0.09,
+                    alpha_height_frac=0.09 * 238 / 387 * 0.75,
+                    ladder=(0.95, 1.0, 1.05),
+                    detect_ncc_threshold=0.50,
+                    provenance_ncc_factor=0.62,
+                    max_saturation=255,
+                )
+            ),
+            BottomRightBoxTextMarkEngine(
+                replace(
+                    _CONFIG,
+                    asset_name="yuanbao_compact_gray_alpha.png",
+                    width_frac=0.28,
+                    height_frac=0.13,
+                    scale_basis="width",
+                    alpha_width_frac=0.06,
+                    alpha_height_frac=0.06 * 238 / 387 * 0.90,
+                    ladder=(0.95, 1.0, 1.05),
+                    detect_frontend="gray",
+                    detect_ncc_threshold=0.50,
+                    provenance_ncc_factor=0.50,
+                    max_saturation=255,
+                )
+            ),
+        )
@@ -14,6 +14,7 @@ ROOT = Path(__file__).resolve().parents[1]
 SKILL = ROOT / "skills" / "remove-ai-watermarks" / "SKILL.md"
 PLUGIN = ROOT / "skills" / ".claude-plugin" / "plugin.json"
 MARKETPLACE = ROOT / ".claude-plugin" / "marketplace.json"
+CODEX_PLUGIN = ROOT / ".codex-plugin" / "plugin.json"
 AGENT_DOC = ROOT / "docs" / "agent-skill.md"
 INSTALL_DOC = ROOT / "docs" / "installation.md"
 INSTALL_REFERENCE = ROOT / "skills" / "remove-ai-watermarks" / "references" / "install.md"
@@ -55,6 +56,16 @@ def test_skill_frontmatter_matches_agent_skills_spec() -> None:
    assert "remove-ai-watermarks" in description


+def test_project_skills_stay_out_of_public_skill_discovery() -> None:
+    # `npx skills add` and skills.sh discover every SKILL.md under .claude/skills;
+    # metadata.internal hides a maintainer-only skill from that listing.
+    project_skills = sorted((ROOT / ".claude" / "skills").glob("*/SKILL.md"))
+    assert project_skills
+    for path in project_skills:
+        frontmatter = _frontmatter(path.read_text(encoding="utf-8"))
+        assert re.search(r"(?m)^metadata:\n  internal: true$", frontmatter), path
+
+
 def test_plugin_and_marketplace_names_agree() -> None:
    plugin = json.loads(PLUGIN.read_text(encoding="utf-8"))
    marketplace = json.loads(MARKETPLACE.read_text(encoding="utf-8"))
@@ -531,15 +542,16 @@ def test_probe_minimum_version_is_a_release_that_exists() -> None:
    assert current >= probe.MIN_CLI_VERSION, "the probe demands a version the project has not released"


-def test_skill_version_agrees_across_all_three_manifests() -> None:
-    """Three files carry this version, and a catalog reads a different one from each."""
+def test_skill_version_agrees_across_every_manifest() -> None:
+    """Four files carry this version, and a catalog reads a different one from each."""
    frontmatter = _frontmatter(SKILL.read_text(encoding="utf-8"))
    match = re.search(r'(?m)^\s+version:\s*"?([0-9][^"\n]*?)"?\s*$', frontmatter)
    assert match is not None, "SKILL.md frontmatter no longer carries metadata.version"
    plugin = json.loads(PLUGIN.read_text(encoding="utf-8"))
    marketplace = json.loads(MARKETPLACE.read_text(encoding="utf-8"))
+    codex = json.loads(CODEX_PLUGIN.read_text(encoding="utf-8"))

-    assert match.group(1) == plugin["version"] == marketplace["metadata"]["version"]
+    assert match.group(1) == plugin["version"] == marketplace["metadata"]["version"] == codex["version"]


 def test_a_bad_invocation_shares_the_no_signal_exit_code(tmp_path: Path) -> None:
@@ -1326,3 +1326,29 @@ def test_unreadable_input_exits_cleanly(runner, tmp_path, name, content, cmd):
    else:
        assert result.exit_code == 1, result.output  # cannot remove a mark from unreadable input
        assert "Error" in result.output
+
+
+@pytest.mark.parametrize(
+    ("name", "content"),
+    [
+        ("empty.png", b""),
+        ("notimage.jpg", b"plain text, not an image at all " * 20),
+        ("truncated.png", b"\x89PNG\r\n\x1a\n" + b"\x00" * 40),
+    ],
+)
+@pytest.mark.parametrize("force", [[], ["--force"]])
+def test_invisible_rejects_unreadable_input_instead_of_reporting_no_signal(runner, tmp_path, name, content, force):
+    """The scrub gate reads undecodable bytes as "no target", so without a decode
+    check `invisible` exited 2 -- the no-signal code -- on a file it never read,
+    while `all` and `batch` exited 1."""
+    bad = tmp_path / name
+    bad.write_bytes(content)
+    mock_cls, mock_engine = _mock_invisible_engine()
+    with (
+        patch("remove_ai_watermarks.invisible_engine.InvisibleEngine", mock_cls),
+        patch("remove_ai_watermarks.invisible_engine.is_available", return_value=True),
+    ):
+        result = runner.invoke(main, ["invisible", str(bad), "-o", str(tmp_path / "out.png"), *force])
+    assert result.exit_code == 1, result.output
+    assert "Failed to read image" in result.output
+    mock_engine.remove_watermark.assert_not_called()
@@ -0,0 +1,75 @@
+"""The Codex plugin manifest, held to the checks a Codex plugin catalog runs on it."""
+
+from __future__ import annotations
+
+import json
+import re
+import subprocess
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[1]
+MANIFEST = ROOT / ".codex-plugin" / "plugin.json"
+CLAUDE_PLUGIN = ROOT / "skills" / ".claude-plugin" / "plugin.json"
+
+
+def _manifest() -> dict:
+    return json.loads(MANIFEST.read_text(encoding="utf-8"))
+
+
+def test_manifest_carries_every_field_the_catalog_validator_requires() -> None:
+    manifest = _manifest()
+    for field in ("name", "version", "description", "repository", "license"):
+        assert str(manifest.get(field, "")).strip(), field
+    interface = manifest["interface"]
+    for field in ("displayName", "shortDescription", "composerIcon"):
+        assert str(interface.get(field, "")).strip(), field
+    assert re.fullmatch(r"[a-z0-9]+(?:-[a-z0-9]+)*", manifest["name"])
+    assert re.fullmatch(r"\d+\.\d+\.\d+", manifest["version"])
+    assert re.fullmatch(r"#[0-9A-Fa-f]{6}", interface["brandColor"])
+
+
+def test_manifest_describes_the_same_plugin_as_the_claude_manifest() -> None:
+    manifest = _manifest()
+    claude = json.loads(CLAUDE_PLUGIN.read_text(encoding="utf-8"))
+    for field in ("name", "description", "repository", "license", "homepage"):
+        assert manifest[field] == claude[field], field
+
+
+def test_every_declared_path_stays_inside_the_repository() -> None:
+    manifest = _manifest()
+    interface = manifest["interface"]
+    paths = [manifest["skills"], interface["composerIcon"], interface["logo"]]
+    for declared in paths:
+        assert declared.startswith("./"), declared
+        resolved = (ROOT / declared).resolve()
+        assert resolved.is_relative_to(ROOT), declared
+        assert resolved.exists(), declared
+    assert (ROOT / manifest["skills"] / "remove-ai-watermarks" / "SKILL.md").is_file()
+
+
+def test_icon_is_a_small_self_contained_svg() -> None:
+    icon = ROOT / _manifest()["interface"]["composerIcon"]
+    content = icon.read_text(encoding="utf-8")
+    assert icon.suffix == ".svg"
+    assert icon.stat().st_size <= 50 * 1024
+    for forbidden in ("data:image/", "base64", "TODO", "PLACEHOLDER"):
+        assert forbidden not in content, forbidden
+
+
+def test_no_tracked_code_matches_the_scanner_dynamic_execution_rule() -> None:
+    """The catalog's scanner reports a high finding for any call spelled like
+    Python's built-in evaluator in its code files, and that rule cannot tell it from
+    the PyTorch module method of the same name. Its scan job is advisory, but the
+    finding still lands on the listing PR, so the code switches modules to inference
+    with the equivalent ``module.train(False)``."""
+    tracked = subprocess.run(
+        ["git", "ls-files", "*.py", "*.js", "*.mjs", "*.cjs", "*.ts", "*.tsx", "*.jsx"],  # noqa: S607
+        cwd=ROOT,
+        check=True,
+        capture_output=True,
+        text=True,
+    ).stdout.split()
+    assert tracked
+    rule = re.compile(r"\beval\s*\(")
+    offenders = [name for name in tracked if rule.search((ROOT / name).read_text(encoding="utf-8", errors="ignore"))]
+    assert offenders == []
@@ -1,20 +1,27 @@
 """Unit tests for how far --cpu-offload travels, per profile.

 The flag sets two residency fields. The face one is read by the shared base and
-reaches every profile; the global one is read by ``QwenZImagePipeline`` alone, so
-on ``chroma-zimage`` and ``sdxl-zimage`` it never touches the stack the user is
-trying to fit. ``WatermarkRemover`` says so before the model load.
+reaches every profile; the global one is read by ``QwenZImagePipeline`` and
+``SdxlZImagePipeline``, so on ``chroma-zimage`` it never touches the stack the user
+is trying to fit. ``WatermarkRemover`` says so before the model load.

 Both halves are exercised with an uninitialized remover, so the core CI matrix
-needs no diffusion dependency, model download, or GPU.
+needs no diffusion dependency, model download, or GPU. The SDXL loader placement is
+driven through stubbed Diffusers constructors and skips without the diffusion extra.
 """

 from __future__ import annotations

 import logging
+from typing import Any
+from unittest.mock import MagicMock

 import pytest

+from remove_ai_watermarks._internal.sdxl_zimage_pipeline import (
+    RESIDENT_SDXL_GLOBAL_MIN_VRAM_GIB,
+    resolve_sdxl_global_residency,
+)
 from remove_ai_watermarks._internal.watermark_profiles import (
    AUTO_PROFILE,
    CHROMA_ZIMAGE_PROFILE,
@@ -29,7 +36,7 @@ from remove_ai_watermarks._internal.watermark_remover import WatermarkRemover
 OFFLOAD_SUPPORT_CASES = (
    (QWEN_ZIMAGE_PROFILE, True),
    (CHROMA_ZIMAGE_PROFILE, False),
-    (SDXL_ZIMAGE_PROFILE, False),
+    (SDXL_ZIMAGE_PROFILE, True),
    # Auto preloads its concrete Qwen fallback before an image vendor is known.
    (AUTO_PROFILE, True),
 )
@@ -69,18 +76,20 @@ class TestGlobalOffloadSupport:


 class TestUnsupportedProfileWarns:
-    @pytest.mark.parametrize("profile", [CHROMA_ZIMAGE_PROFILE, SDXL_ZIMAGE_PROFILE])
-    def test_it_names_the_profile_and_the_flag(self, profile: str, caplog: pytest.LogCaptureFixture) -> None:
+    def test_it_names_the_profile_and_the_flag(self, caplog: pytest.LogCaptureFixture) -> None:
        with caplog.at_level(logging.WARNING):
-            _remover(profile, cpu_offload=True)._warn_if_global_offload_unsupported()
+            _remover(CHROMA_ZIMAGE_PROFILE, cpu_offload=True)._warn_if_global_offload_unsupported()
        assert len(caplog.records) == 1
        message = caplog.records[0].getMessage()
        assert "--cpu-offload" in message
-        assert profile in message
+        assert CHROMA_ZIMAGE_PROFILE in message

-    def test_qwen_zimage_is_quiet(self, caplog: pytest.LogCaptureFixture) -> None:
+    @pytest.mark.parametrize("profile", [QWEN_ZIMAGE_PROFILE, SDXL_ZIMAGE_PROFILE])
+    def test_a_profile_that_streams_its_global_stack_is_quiet(
+        self, profile: str, caplog: pytest.LogCaptureFixture
+    ) -> None:
        with caplog.at_level(logging.WARNING):
-            _remover(QWEN_ZIMAGE_PROFILE, cpu_offload=True)._warn_if_global_offload_unsupported()
+            _remover(profile, cpu_offload=True)._warn_if_global_offload_unsupported()
        assert caplog.records == []

    def test_a_run_that_did_not_ask_for_offload_is_quiet(self, caplog: pytest.LogCaptureFixture) -> None:
@@ -132,3 +141,118 @@ class TestUnsupportedProfileWarns:
        with pytest.raises(ImportError):
            remover._load_qwen_zimage_pipeline()
        assert order == ["warned", "loaded"]
+
+
+class TestSdxlGlobalResidency:
+    @pytest.mark.parametrize(
+        ("requested", "total_memory_gib", "expected"),
+        [
+            # Unset: the card decides, and an 8 GiB card cannot hold the stack.
+            (None, 7.6, False),
+            (None, RESIDENT_SDXL_GLOBAL_MIN_VRAM_GIB, True),
+            (None, 23.5, True),
+            # An unreadable card reports 0.0 and must stream rather than spill.
+            (None, 0.0, False),
+            # Explicit requests win in both directions.
+            (False, 79.2, False),
+            (True, 7.6, True),
+        ],
+    )
+    def test_request_wins_and_the_card_decides_otherwise(
+        self, requested: bool | None, total_memory_gib: float, expected: bool
+    ) -> None:
+        assert resolve_sdxl_global_residency(requested, total_memory_gib=total_memory_gib) is expected
+
+
+class _RecordingSdxlPipe:
+    """Stands in for the Diffusers SDXL pipeline and records placement calls in order."""
+
+    def __init__(self, calls: list[Any]) -> None:
+        self.calls = calls
+        self.scheduler = MagicMock()
+
+    def to(self, device: str) -> _RecordingSdxlPipe:
+        self.calls.append(("to", device))
+        return self
+
+    def load_lora_weights(self, _path: str) -> None:
+        self.calls.append(("load_lora_weights",))
+
+    def fuse_lora(self) -> None:
+        self.calls.append(("fuse_lora",))
+
+    def enable_sequential_cpu_offload(self, *, device: str, gpu_id: int | None = None) -> None:
+        self.calls.append(("enable_sequential_cpu_offload", device, gpu_id))
+
+
+class TestSdxlGlobalPlacement:
+    @staticmethod
+    def _load(
+        monkeypatch: pytest.MonkeyPatch,
+        *,
+        requested: bool | None,
+        total_memory_gib: float,
+        current_device: int = 0,
+    ) -> list[Any]:
+        diffusers = pytest.importorskip("diffusers")
+        torch = pytest.importorskip("torch")
+        monkeypatch.setattr(torch.cuda, "current_device", lambda: current_device)
+        from remove_ai_watermarks._internal.sdxl_zimage_pipeline import SdxlZImagePipeline
+
+        calls: list[Any] = []
+        monkeypatch.setattr(diffusers.ControlNetModel, "from_pretrained", lambda *a, **k: MagicMock())
+        monkeypatch.setattr(diffusers.AutoencoderKL, "from_pretrained", lambda *a, **k: MagicMock())
+        monkeypatch.setattr(
+            diffusers.StableDiffusionXLControlNetImg2ImgPipeline,
+            "from_pretrained",
+            lambda *a, **k: _RecordingSdxlPipe(calls),
+        )
+        monkeypatch.setattr("huggingface_hub.hf_hub_download", lambda *a, **k: "lora.safetensors")
+        monkeypatch.setattr(diffusers.EulerDiscreteScheduler, "from_config", lambda *a, **k: MagicMock())
+
+        pipeline = SdxlZImagePipeline(device="cuda", torch_dtype=None, keep_global_models_on_device=requested)
+        monkeypatch.setattr(type(pipeline), "_require_cuda", lambda self: None)
+        monkeypatch.setattr(type(pipeline), "_total_vram_gib", lambda self: total_memory_gib)
+        pipeline._load_global()
+        return calls
+
+    def test_a_small_card_streams_the_fused_stack_instead_of_moving_it(self, monkeypatch: pytest.MonkeyPatch) -> None:
+        calls = self._load(monkeypatch, requested=None, total_memory_gib=7.6)
+        # The offload hooks stream their own copy of each weight, so they must follow the fuse.
+        assert calls == [("load_lora_weights",), ("fuse_lora",), ("enable_sequential_cpu_offload", "cuda", 0)]
+
+    def test_a_large_card_keeps_the_stack_resident_as_before(self, monkeypatch: pytest.MonkeyPatch) -> None:
+        calls = self._load(monkeypatch, requested=None, total_memory_gib=23.5)
+        assert calls == [("to", "cuda"), ("load_lora_weights",), ("fuse_lora",)]
+
+    def test_cpu_offload_streams_even_on_a_large_card(self, monkeypatch: pytest.MonkeyPatch) -> None:
+        calls = self._load(monkeypatch, requested=False, total_memory_gib=79.2)
+        assert ("to", "cuda") not in calls
+        assert calls[-1] == ("enable_sequential_cpu_offload", "cuda", 0)
+
+    @pytest.mark.parametrize("current_device", [1, 2])
+    @pytest.mark.parametrize(("requested", "total_memory_gib"), [(None, 7.6), (False, 79.2)])
+    def test_streaming_uses_the_current_cuda_device(
+        self, monkeypatch: pytest.MonkeyPatch, current_device: int, requested: bool | None, total_memory_gib: float
+    ) -> None:
+        calls = self._load(
+            monkeypatch,
+            requested=requested,
+            total_memory_gib=total_memory_gib,
+            current_device=current_device,
+        )
+        assert ("to", "cuda") not in calls
+        assert calls[-1] == ("enable_sequential_cpu_offload", "cuda", current_device)
+
+    def test_cpu_offload_reaches_the_sdxl_stack(self, monkeypatch: pytest.MonkeyPatch) -> None:
+        from remove_ai_watermarks._internal import sdxl_zimage_pipeline
+
+        captured: dict[str, object] = {}
+
+        class Recorder:
+            def __init__(self, **kwargs: object) -> None:
+                captured.update(kwargs)
+
+        monkeypatch.setattr(sdxl_zimage_pipeline, "SdxlZImagePipeline", Recorder)
+        _remover(SDXL_ZIMAGE_PROFILE, cpu_offload=True)._load_qwen_zimage_pipeline()
+        assert captured["keep_global_models_on_device"] is False
Author	SHA1	Message	Date
Victor KuznetsovandClaude Opus 5.5	5677c22a6b	Release 0.46.0 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>	2026-10-17 14:31:59 -07:00
Victor KuznetsovandClaude Opus 5.5	a4eb16ad77	Build the expected MCP output path with the platform separator Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>	2026-10-17 14:27:35 -07:00
Victor KuznetsovandClaude Opus 5.5	06c0cb1bbe	Add Codex plugin manifest, stdlib MCP server, SHA-pinned Actions - .codex-plugin/plugin.json and assets/icon.svg package the existing skill tree for Codex plugin catalogs; skill version 1.0.12 across all manifests - `remove-ai-watermarks mcp` serves identify, remove and batch over stdio, one CLI command per tool call, standard library only; server.json and glama.json for the MCP registry and Glama - Pin every GitHub Action to a commit SHA - Call module.train(False) instead of module.eval(), which plugin scanners read as dynamic code execution Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>	2026-10-17 14:13:04 -07:00
Victor Kuznetsov	db006f034f	Merge pull request #123 from wiltodelta/docs/readme-first-screen Move the before/after demo and install lines to the README first screen	2026-10-17 12:55:01 -07:00
Victor KuznetsovandClaude Opus 5.5	604a8a8f52	Move the before/after demo and install lines to the README first screen Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>	2026-10-17 12:30:46 -07:00
Victor KuznetsovandClaude Opus 5.5	19ae5d1b0a	Strip app-export and Kling SEI provenance from MP4, refresh watermark docs video metadata --remove now blanks app-export JSON in keyed and keyless metadata items (Dreamina writes it twice), which it previously left in place while exiting 1. Kling writes an H.264 user_data_unregistered SEI message (a fixed UUID plus "kling-ai") into the first video sample. get_ai_metadata reports it as kling_sei, identify_video names Kling from it when no TC260 label does, and the streaming stripper blanks every copy inside mdat during its existing copy pass. Lengths stay, the SEI stays well-formed, and ffmpeg framemd5 hashes of every frame match before and after on both tracked Kling fixtures. Across 255 local files only the seven Kling videos changed verdict. invisible now exits 1 on an undecodable source instead of reporting the no-signal code for a file it never read. Docs: a per-command "what each path leaves behind" table, the October literature sweep, Suno's announced audio watermark, Anthropic's SynthID-Text-based Claude watermark, and an independent TrustMark false-positive benchmark. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>	2026-10-17 12:30:01 -07:00
Victor Kuznetsov	21f4621eb7	Add SynthID Detector surface, drop extra web confirmation	2026-10-17 11:47:53 -07:00
Victor Kuznetsov	f1d4d4afba	Merge pull request #121 from dukagjina/fix/preserve-invisible-transparency Preserve alpha in invisible image processing	2026-10-16 18:06:56 -07:00
Victor Kuznetsov	2430689f96	Merge branch 'main' into fix/preserve-invisible-transparency	2026-10-16 12:36:43 -07:00
Victor KuznetsovandClaude Opus 5.5	cf695de57f	Refresh the C2PA soft-binding snapshot and fix Ruff ICN001 The upstream registry gained be.immostory.videoseal.1 at revision d5960248, so `sync_c2pa_soft_bindings.py --check` failed lint on main and on every pull request. That step runs ahead of Ruff, which hid an ICN001 error in scripts/_corpus_scan.py; import `date` directly, as the sibling scripts do. Document the check's live-upstream behavior next to the refresh command. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>	2026-10-16 12:15:03 -07:00
dukagjina	a054185fa5	Preserve source transparency during invisible image processing	2026-10-16 02:04:54 -04:00
Victor Kuznetsov	3e20bb0b92	Merge pull request #120 from wiltodelta/dependabot/uv/minor-and-patch-7fed723961	2026-10-14 22:27:16 -07:00
dependabot[bot]	0b4170ae98	Bump the minor-and-patch group with 6 updates Bumps the minor-and-patch group with 6 updates: \| Package \| From \| To \| \| --- \| --- \| --- \| \| [pillow-heif](https://github.com/bigcat88/pillow_heif) \| `1.8.0` \| `1.8.1` \| \| [diffusers](https://github.com/huggingface/diffusers) \| `0.40.0` \| `0.41.0` \| \| [diffsynth](https://github.com/modelscope/DiffSynth-Studio) \| `2.1.8` \| `2.1.9` \| \| [onnxruntime](https://github.com/microsoft/onnxruntime) \| `1.30.0` \| `1.31.0` \| \| [openai](https://github.com/openai/openai-python) \| `3.24.0` \| `3.29.0` \| \| [ruff](https://github.com/astral-sh/ruff) \| `0.16.10` \| `0.17.0` \| Updates `pillow-heif` from 1.8.0 to 1.8.1 - [Release notes](https://github.com/bigcat88/pillow_heif/releases) - [Changelog](https://github.com/bigcat88/pillow_heif/blob/master/CHANGELOG.md) - [Commits](https://github.com/bigcat88/pillow_heif/compare/v1.8.0...v1.8.1) Updates `diffusers` from 0.40.0 to 0.41.0 - [Release notes](https://github.com/huggingface/diffusers/releases) - [Commits](https://github.com/huggingface/diffusers/compare/v0.40.0...v0.41.0) Updates `diffsynth` from 2.1.8 to 2.1.9 - [Release notes](https://github.com/modelscope/DiffSynth-Studio/releases) - [Commits](https://github.com/modelscope/DiffSynth-Studio/commits) Updates `onnxruntime` from 1.30.0 to 1.31.0 - [Release notes](https://github.com/microsoft/onnxruntime/releases) - [Changelog](https://github.com/microsoft/onnxruntime/blob/main/docs/ReleaseNotesWorkflow.md) - [Commits](https://github.com/microsoft/onnxruntime/compare/v1.30.0...v1.31.0) Updates `openai` from 3.24.0 to 3.29.0 - [Release notes](https://github.com/openai/openai-python/releases) - [Changelog](https://github.com/openai/openai-python/blob/main/CHANGELOG.md) - [Commits](https://github.com/openai/openai-python/compare/v3.24.0...v3.29.0) Updates `ruff` from 0.16.10 to 0.17.0 - [Release notes](https://github.com/astral-sh/ruff/releases) - [Changelog](https://github.com/astral-sh/ruff/blob/main/CHANGELOG.md) - [Commits](https://github.com/astral-sh/ruff/compare/0.16.10...0.17.0) --- updated-dependencies: - dependency-name: pillow-heif dependency-version: 1.8.1 dependency-type: direct:production update-type: version-update:semver-patch dependency-group: minor-and-patch - dependency-name: diffusers dependency-version: 0.41.0 dependency-type: direct:production update-type: version-update:semver-minor dependency-group: minor-and-patch - dependency-name: diffsynth dependency-version: 2.1.9 dependency-type: direct:production update-type: version-update:semver-patch dependency-group: minor-and-patch - dependency-name: onnxruntime dependency-version: 1.31.0 dependency-type: direct:production update-type: version-update:semver-minor dependency-group: minor-and-patch - dependency-name: openai dependency-version: 3.29.0 dependency-type: direct:production update-type: version-update:semver-minor dependency-group: minor-and-patch - dependency-name: ruff dependency-version: 0.17.0 dependency-type: direct:production update-type: version-update:semver-minor dependency-group: minor-and-patch ... Signed-off-by: dependabot[bot] <support@github.com>	2026-10-14 20:05:51 +00:00
Victor Kuznetsov	4241a01659	Release 0.45.1	2026-10-08 16:42:57 -07:00
Victor Kuznetsov	c37378b347	Merge pull request #117 from dichiara19/fix/sdxl-zimage-global-offload Stream the sdxl-zimage global stack on cards below 12 GiB	2026-10-08 16:34:43 -07:00
Victor Kuznetsov	8cb99b73b2	Fix SDXL offload placement on the current CUDA device	2026-10-08 16:17:11 -07:00
Giuseppe G. Di Chiara	c45e6db3e4	Stream the sdxl-zimage global stack on cards below 12 GiB The SDXL global stack (fp16 UNet, Canny ControlNet, both text encoders, VAE) is ~8.8 GiB and was always moved to CUDA with .to(device), so it could not fit an 8 GiB card and --cpu-offload never reached it. Under the Windows display driver, WSL2 included, the overflow spilled into shared system memory and the four-step pass ran for over ten minutes instead of failing. Residency now follows the same rule as the Qwen stack: an explicit request wins, otherwise the card's total VRAM decides against RESIDENT_SDXL_GLOBAL_MIN_VRAM_GIB. Below it the Lightning LoRA is fused on the CPU and the stack is streamed with enable_sequential_cpu_offload, the fastest placement that completed on every run of an 8 GiB card with 16 GiB of system RAM. sdxl-zimage joins GLOBAL_OFFLOAD_PROFILES, so the --cpu-offload warning now fires for chroma-zimage only.	2026-10-08 23:27:55 +02:00
Victor Kuznetsov	37fb080c74	Merge pull request #116 from wiltodelta/dependabot/uv/minor-and-patch-931ef6a6bd	2026-10-08 10:30:09 -07:00
dependabot[bot]	954298aa14	Bump the minor-and-patch group with 3 updates Bumps the minor-and-patch group with 3 updates: [python-dotenv](https://github.com/theskumar/python-dotenv), [openai](https://github.com/openai/openai-python) and [ruff](https://github.com/astral-sh/ruff). Updates `python-dotenv` from 1.2.3 to 1.2.4 - [Release notes](https://github.com/theskumar/python-dotenv/releases) - [Changelog](https://github.com/theskumar/python-dotenv/blob/main/CHANGELOG.md) - [Commits](https://github.com/theskumar/python-dotenv/compare/v1.2.3...v1.2.4) Updates `openai` from 3.22.1 to 3.24.0 - [Release notes](https://github.com/openai/openai-python/releases) - [Changelog](https://github.com/openai/openai-python/blob/main/CHANGELOG.md) - [Commits](https://github.com/openai/openai-python/compare/v3.22.1...v3.24.0) Updates `ruff` from 0.16.9 to 0.16.10 - [Release notes](https://github.com/astral-sh/ruff/releases) - [Changelog](https://github.com/astral-sh/ruff/blob/main/CHANGELOG.md) - [Commits](https://github.com/astral-sh/ruff/compare/0.16.9...0.16.10) --- updated-dependencies: - dependency-name: python-dotenv dependency-version: 1.2.4 dependency-type: direct:production update-type: version-update:semver-patch dependency-group: minor-and-patch - dependency-name: openai dependency-version: 3.24.0 dependency-type: direct:production update-type: version-update:semver-minor dependency-group: minor-and-patch - dependency-name: ruff dependency-version: 0.16.10 dependency-type: direct:production update-type: version-update:semver-patch dependency-group: minor-and-patch ... Signed-off-by: dependabot[bot] <support@github.com>	2026-10-07 08:48:45 +00:00
Victor KuznetsovandClaude Opus 5.5	10c7775990	Hide maintainer skills from public skill discovery The skills CLI discovers every SKILL.md under .claude/skills, so provider-oracles and signal-discovery were offered by `npx skills add` and listed on skills.sh beside the published skill. Mark both metadata.internal, the CLI's documented hide flag, and fail the suite on any project skill that lacks it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>	2026-10-05 20:04:18 -07:00
Victor KuznetsovandClaude Opus 5.5	3f6447b23d	Record how to run the gate from an agent session The full gate outlasts the agent's 10-minute foreground shell cap, so it must run detached with its exit code written to a file. A stray run is stopped by its own PID, because a pattern kill also ends other sessions' suites. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>	2026-10-05 18:25:14 -07:00
Victor KuznetsovandClaude Opus 5.5	cc63bfd7bf	Release 0.45.0 Bump the package to 0.45.0 and the agent skill to 1.0.11. The skill now names the dola and workbuddy image mark keys, which 0.44.0 rejects, so the probe's MIN_CLI_VERSION rises to 0.45.0 and both keys carry a "since CLI" note that the floor test reads. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>	2026-10-05 17:45:18 -07:00
Victor Kuznetsov	898f8be789	Add AI disclosure layouts, tighten detection and provenance warnings	2026-10-04 11:03:37 -07:00
Victor KuznetsovandClaude Opus 5.5	72752feb33	Scan uv.lock in CI and add a security policy Add a CI job running the maintain.sh uv-secure scan, pinned to the gate by a parity test, and a SECURITY.md with private reporting and dependency remediation. Refresh the security-gate block record for accelerate 1.15.0. Closes #114 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>	2026-10-02 10:33:50 -07:00
Victor KuznetsovandClaude Opus 5.5	70adddf960	Record why Qwen-Image-2.1 is not a regeneration profile Open weights remove the hosted-only objection to the 2.x line, but the Diffusers pipeline has no strength or latent noise-mask input and the Qwen Research License is non-commercial (issue #111). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>	2026-10-02 10:33:49 -07:00
Victor Kuznetsov	01a2665181	Reconcile research follow-up status	2026-10-01 22:10:00 -07:00