Two AI-provenance metadata types mined from the retained corpus that
identify previously read as no-signal:
- Dreamina (ByteDance's international Jimeng brand) signs C2PA as
"Bytedance Pte. Ltd." with a "Dreamina/x.y" claim generator and NO
digitalSourceType, so the generator name is the only AI signal. Add a
C2paAiVendor row with a new asserts_ai flag (identity-AI: presence
asserts AI without trainedAlgorithmicMedia) plus the derived
C2PA_IDENTITY_AI_ORGS view, folded into identify's c2pa_is_ai. Keyed on
the Dreamina generator token, not the "Bytedance Pte" issuer, so non-AI
CapCut edits signed by the same entity stay unattributed. 7/7 corpus
files now attribute to ByteDance.
- Tencent Cloud's TC260 AIGC variant uses a ServiceProvider/ServiceUser
schema (vs the producer-side ContentProducer schema), embedded in EXIF
ImageDescription; add those field names to _TC260_FIELDS so the generic
{"AIGC":{...}} gate accepts it. 11/11 corpus files now flagged.
Test-first: reproducing tests in test_identify.py / test_metadata.py.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
main landed #58 (pill-gate fix, superseded by this branch's localize->fill
rewrite) and #57 (deps bump). Resolved the 6 code/test/doc conflicts by keeping
this branch's post-rewrite versions; the deps bump auto-merged into
uv.lock/pyproject. Full gate green after resolution: ruff, pyright 0, 730 passed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The FP gate demotes a low-gradient match, but a real FAINT sparkle also has soft
edges, so metadata-stripped faint sparkles were dropped. Keep a low-grad match
that is a strong (conf >= 0.52), bright, near-WHITE-core sparkle: a real sparkle
core is white, a clean bright corner that shape-matches (sky/sun) is colored
(_core_saturation). Recovers ~14/20 stripped faint sparkles under the DEFAULT
strict/auto (no metadata, no flag) at ~1.25% clean false-fire (baseline 0.55%);
the ~0.51-scoring bright-background FPs stay demoted (below 0.52).
A learned classifier on the same features measured WORSE than the tuned gate
(tier-1: MLP 86.7% recall vs the gate's 90.8% at equal false-fire), so the
heuristic stays; a patch-CNN with richer features is roadmapped P2 with low
expected value -- the precision/recall wall is fundamental (deep-research +
tier-1 both confirm it).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Deep-research 2026-07-10 (adversarially verified): the Gemini sparkle is
tier-gated (visible on Free/Pro, absent on Ultra/AI-Studio/API; no official
visible-mark detector or published glyph spec); the faint-visible-mark
precision/recall wall is fundamental (learned CNN front-end does not cleanly
separate true/false, arXiv:1705.08593 refuted); learned detectors need large
synthetic-composite datasets + carry off-distribution risk; landscape adds
Meta bottom-left + Samsung star-icon variants; China GB 45438-2025 is the
strongest visible-mark mandate.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Full-dataset validation of reverse-alpha (v0.12.1) vs the current localize->fill:
doubao/jimeng identical (100% coverage + clearance across all backends); gemini
strict coverage a few points below reverse-alpha (the FP tightening), every
missed mark recovered under assume_ai, clearance ~98% both, no outside-box
damage. Clearance is fill-independent (cv2/MI-GAN/LaMa all strip the mark shape);
the difference is visual fill quality on textured/structured backgrounds -- LaMa
best, MI-GAN can ghost/hallucinate, cv2 smears -- which motivates auto = LaMa >
MI-GAN > cv2. Added to module-internals, known-limitations, and the CLAUDE.md
compact list.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The auto backend now resolves best-first: LaMa (highest quality, recovers the
textured/structured backgrounds the classical fill smears) > MI-GAN > cv2. Both
learned backends share the same onnxruntime availability check, so auto cannot
tell them apart and always prefers the better one; a memory-tight deployment
that cannot afford LaMa's ~4.7 GB peak pins MI-GAN explicitly via
`--backend migan` / `backend="migan"` (the deployment's call, not the library's).
cv2 stays the no-deps floor and now emits a one-time quality warning when auto
falls back to it, since it smears texture/structure.
Motivated by a v0.12.1 reverse-alpha vs 0.14 localize->fill head-to-head:
reverse-alpha recovered structured backgrounds more cleanly than any inpaint;
LaMa closes most of that gap, MI-GAN can ghost/hallucinate, cv2 is weakest.
doubao/jimeng removal is identical between versions; gemini strict coverage is
4pp lower (all recovered via assume_ai) with cleaner clearance and no
outside-box damage.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Candidate carries only the fields the arbiter reads (key, label,
detected_strict, detected_relaxed, features); location/region/confidence were
vestigial from the removed best_auto_mark max-by-confidence path.
- resolve_backend returns preferred_inpaint_backend() directly (typed Literal)
instead of an identity ternary.
- colour/normalise/behaviour -> US spelling across code comments and docs.
No behavior change.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Resolve 10 code-review findings on the v0.14.0 localize->fill path, several
release-blocking:
- gemini: build the removal mask from the decision's provenance-aware region
instead of a strict internal re-detect. A relaxed/assume_ai sparkle was
re-demoted by the FP gate into a None mask and reported removed while left in
the image; this also drops the redundant double-detect.
- registry: report a mark removed only when a fill actually happened (remove()
returns a None region for an empty mask), so a no-op is never claimed.
- api/cli: add write_noop so the CLI `visible` no-mark path writes nothing and
cannot clobber a pre-existing -o file (was write-then-unlink -> data loss);
create output.parent; skip the same-file copy (SameFileError on in-place).
- cli: catch the missing migan/lama backend RuntimeError on the visible/all
paths (matches `erase`); route the single-mark relaxation through the shared
resolve_relax instead of an inline copy.
- metadata: keep_standard=False no longer takes the AI-only lossless JPEG
short-circuit (it left standard metadata); defer a malformed-marker JPEG to
the PIL fallback instead of reporting a partial strip as complete.
- invisible: register the HEIF opener before Image.open (HEIC --force) and
RGB-convert before the PNG temp (CMYK JPEG).
- pill: normalize via to_bgr so a 4-channel BGRA array cannot crash cvtColor.
Regression tests for each; docs synced (resolve_relax, write_noop,
best_auto_mark -> detect_marks).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Replace reverse-alpha removal with localize -> fill (template-free mask + one
shared cv2/MI-GAN/big-LaMa fill) for every mark; drops the colour-shift / dark-pit
failure modes, version-robust to a moved or re-rendered mark
- Separate perception/decision/action: engines report Candidates, a pure
decide(candidates, Context) arbiter owns all policy (sensitivity + provenance +
pill gate), remove_auto_marks orchestrates -- behavior-preserving (corpus 46/46/92)
- Three orthogonal knobs replace --method: --backend cv2|migan|lama,
--sensitivity auto|strict|assume-ai, provenance (auto from metadata)
- Add high-level api.remove_visible / visible_provenance (lazy top-level re-export);
visible --mark auto delegates to it so CLI and library share ONE path
- Read+write HEIC/AVIF on the pixel path via pillow-heif; imwrite preserves the input
format at max quality (JPEG q100/4:4:4); a no-op copies the original bytes verbatim
- Lossless byte-level JPEG metadata strip (no DCT re-encode); consolidate the two
remove_ai_metadata into one, delete legacy noai/cleaner + best_auto_mark
- Bump 0.13.0 -> 0.14.0
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Verified 0.13.0 pill removal on a 32k real-upload corpus. The metadata-OR-wordmark
gate was only ~1/3 precise: TC260 metadata confirms Jimeng-class provenance, not pill
presence, so the weak edge-NCC detector's false fires (textured ceilings/walls, where
inpaint visibly smears) were admitted whenever metadata was present.
Split into two arms (_keep_pill): the reliable bottom-right wordmark (~94% precise,
survives metadata stripping) removes the pill unrestricted; the metadata-only arm
removes it ONLY when the top-left footprint is flat enough for an invisible inpaint
(PillEngine.footprint_is_flat, median-Sobel <= _FLAT_TEXTURE_MAX). Keeps real
flat-scene pills and harmless flat false fires; leaves the damaging textured false
fires untouched. Corpus: 270 -> 118 removals, ~90 true preserved, damaging FP -> ~0.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Verified 0.13.0 pill removal on a 32k real-upload corpus. The metadata-OR-wordmark
gate was only ~1/3 precise: TC260 metadata confirms Jimeng-class provenance, not pill
presence, so the weak edge-NCC detector's false fires (textured ceilings/walls, where
inpaint visibly smears) were admitted whenever metadata was present.
Split into two arms (_keep_pill): the reliable bottom-right wordmark (~94% precise,
survives metadata stripping) removes the pill unrestricted; the metadata-only arm
removes it ONLY when the top-left footprint is flat enough for an invisible inpaint
(PillEngine.footprint_is_flat, median-Sobel <= _FLAT_TEXTURE_MAX). Keeps real
flat-scene pills and harmless flat false fires; leaves the damaging textured false
fires untouched. Corpus: 270 -> 118 removals, ~90 true preserved, damaging FP -> ~0.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Add the Jimeng-basic top-left "AI生成" pill as a CAPTURE-LESS mark
(pill_engine.py): synthetic-silhouette edge-NCC detect + inpaint-only removal.
Gated in remove_auto_marks: kept only when Jimeng is confirmed (TC260 metadata
OR the bottom-right "★ 即梦AI" wordmark fired -- the wordmark keeps recall on
metadata-STRIPPED uploads) AND Doubao did not fire.
- Add an inpaint-fallback removal path + MI-GAN ONNX backend (migan extra, MIT,
~28 MB / ~1 GB peak -- droplet-friendly) alongside big-LaMa. New
--method auto|reverse-alpha|inpaint (shared across visible/all/batch) and
erase --backend migan; footprint_mask on each engine.
- auto is deterministic: reverse-alpha for capture marks (recovers exact pixels,
lighter -- measured cleaner than MI-GAN on structured backgrounds) and inpaint
only for the capture-less pill.
- --mark auto now removes EVERY detected mark in one pass (remove_auto_marks),
so a Jimeng-basic image's top-left pill AND bottom-right wordmark both clear.
- Bump 0.12.1 -> 0.13.0.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Bright-background photos/renders and a tiny app icon were flagged as
AI-generated by the visible detectors. Two failure modes:
- Gemini sparkle on a bright background (snow+sky photo, white product
render) scored ~0.51. The FP gate only demoted on a low core-ring
brightness margin, which a bright background makes high. Add a gradient
floor (_SPARKLE_FP_GRAD 0.55): a real sparkle is a crisp star (grad
~0.97-1.0), a smooth luminance blob that NCC-matches the diamond is not
(the two FPs measured grad 0.105 / 0.463). The OR is a strict superset
of the old margin-only demotion, so it cannot regress dark/mid (kept by
margin) or white-bg (kept by confidence) real sparkles.
- A 48x48 geometric icon matched the Doubao/Jimeng CJK silhouette at
0.41/0.47 NCC. Purely a small-size artifact (the same icon at >=256px
collapses to ~0.06-0.10). Guard text-mark detection below a 200px short
side (_MIN_DETECT_SHORT_SIDE); real marks ship on full-resolution
renders (smallest captured sample 1086px).
Corpus re-sweep flips only OpenAI content and already-cleaned outputs,
all sub-0.5, so no provenance verdict changes. Add synthetic regression
fixtures for both modes; docs/module-internals.md updated.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
InvisibleEngine loads SDXL/ControlNet in fp16 on CUDA/XPU but called from_pretrained
without variant="fp16", so it read the full fp32 weight files (~7 GB) and downcast in
memory. _load_from_pretrained now passes variant="fp16" when torch_dtype is float16,
reading the half-precision files (~3.5 GB) instead - roughly halving the cold-start
weight read + host->device transfer (a phase-timed Modal run measured weight load as
~half of the ~25s cold start). Falls back to the default weights when a checkpoint ships
no fp16 variant (a custom --model), so the worst case is the prior behavior. fp32
(cpu/mps) and bf16 (qwen) never request the variant.
Tests: TestFp16WeightVariant (variant requested on fp16, fallback on missing, never on
fp32).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Per user decision 2026-06-22: synthetic font-rendered alpha reconstruction is
rejected as below the quality bar; the reverse-alpha alpha map must be solved
from real controlled flat captures (visible_alpha_solve.py). Meta AI, more
Samsung locales, and any Grok visible mark are parked until captures exist.
Future sessions must not propose synthetic or derive assets from the corpus.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Mined from the retained corpus 2026-06-22 (open-world EXIF/PNG-text/XMP scan,
minus the registry): three AI image generators that stamp a plain generator
name and no C2PA, so identify read them as no-signal -- and under the P0#5
no-signal skip would have skipped the scrub.
- NovelAI (anime SD): PNG tEXt Software/Source/Title. exif_generator now reads
PNG text chunks (via img.info), not only EXIF/XMP.
- Reve (reve.com): EXIF Software / XMP CreatorTool. Token is the full
"reve.com", not bare "reve" (would false-fire on "forever"/"reverie").
- Aphrodite AI: EXIF Make / Software.
Detection/removal parity: NovelAI stamps an AI-shaped VALUE under a non-AI KEY
(Title/Source), which _is_ai_key alone keeps. New _is_ai_value drops a text
chunk by value-token match on removal, mirroring exif_generator -- else the
cleaned file still read as NovelAI (verified on a real corpus file).
Tests: TestExifGenerator gains NovelAI PNG-text, Reve, Reve-not-overmatched,
Aphrodite, and a NovelAI detect/remove parity regression. Docs synced
(module-internals, watermarking-landscape, CLAUDE.md).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Regenerating pixels removes SynthID / open watermarks but degrades a real
photo, so running it on a clean image is the dominant paid score-0 cause on
no-watermark uploads. Gate invisible/all/batch on identify.has_invisible_target:
when no invisible AI signal is locally detectable and --force is unset, skip the
regeneration. Per-command semantics:
- invisible: write no output, exit EXIT_NO_INVISIBLE_SIGNAL (2)
- all: skip step 2 but keep visible-removed pixels + strip metadata, exit 0
- batch: skip the scrub; copy the input through in invisible mode
A skip never claims the image is clean (a pixel SynthID is undetectable once its
metadata proxy is gone); the message says so and routes to --force. The gate
fails safe (a detector error runs the removal).
has_invisible_target wraps identify(check_visible=False, check_invisible=True)
and returns the new ProvenanceReport.ai_from_metadata field (the confidence==high
union), so the raiw.cc worker can reuse the same gate. Gate placed before engine
construction so the skip path is cheap; shared via cli._should_skip_invisible_scrub.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The qwen oracle floors are certified, not pending. Near-threshold scrub is
seed-non-deterministic, but the prod path pins one fixed seed, so a certified
floor reproduces run-to-run -- pinning the seed is the release gate, not a seed
sweep. Reword the docstring so it stops implying an open seed-repeat gate.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Replace the `data/spaces/originals/` path with a generic "local corpus of
pristine originals" so the committed public doc carries no reference to the
local working-data pull (the data itself is gitignored). The analysis scripts'
default paths are left untouched (operational tooling, no content/provenance).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
21k user-pull images under data/spaces/ were untracked but not ignored, so a
stray `git add -A` could have committed them. Add the ignore entry alongside the
other local-data paths; the dir stays local analysis only, never a committed corpus.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- watermark_remover: _build_qwen_kwargs now passes explicit height/width (via
_qwen_target_size, floored to /16). Without it QwenImageImg2ImgPipeline defaults to
1024x1024 and silently squishes non-square inputs, distorting the scene and garbling text.
- watermark_profiles: resolve_strength gains a `pipeline` arg + a Qwen strength ladder
(_QWEN_VENDOR_STRENGTH, Gemini 0.25), so `--pipeline qwen` gets its certified floor
automatically; retires the manual "pass --strength 0.25 for Gemini on qwen" workaround.
- fidelity_metrics: replace per-face nearest matching (collided on multi-face images when a
variant dropped a face, corrupting the identity metric) with a collision-free one-to-one
assignment (assign_faces_one_to_one). lapvar/LPIPS were always bbox-anchored and immune.
Regression-guarded by tests/test_fidelity_matching.py.
- docs: record the measured outcomes of the qwen-improvement arc. The Qwen ControlNet
face-fix is CLOSED (no permissive Qwen detail/tile ControlNet exists; canny carries edges,
not skin grain). The `--pipeline auto` router + faces+text mixed dual-pass were prototyped
and DROPPED (controlnet wins faces AND display text: abba CER 0.114 vs qwen 0.379).
Z-Image-Turbo was tried and dropped (same regeneration limits). qwen stays a manual opt-in;
controlnet is the default for everything.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Cited deep-research report (22 sources, 3-vote adversarial verification, 5 refuted)
behind the "ship qwen as-is or improve first?" decision. Verdict: shippable now as
an opt-in text lane; strongest improvement lead is adding a Qwen-Image ControlNet
(InstantX / DiffSynth, Apache-2.0, diffusers QwenImageControlNetPipeline) for face/
skin structure; Z-Image-Turbo (6B, Apache-2.0) is the best cheaper text-preserving
substitute. No improvement has measured face-fidelity at our scrub floors yet --
validate with scripts/fidelity_metrics.py first. Linked from known-limitations.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Update CLAUDE.md and docs/module-internals.md for: ProvenanceReport.ai_source_kind
(generated vs enhanced) and the shared GEMINI_SPARKLE_TRUST_CONF; the text-mark
over-subtraction guard; noai/tiling.feather_region_composite + the region-targeted
WatermarkRemover.remove_watermark(region=) path; the new C2PA vendor rows (Volcano
Engine Chinese legal name, ElevenLabs) and the documented TikTok/PixelBin
exclusion. Record the rejected gemini-gate-lowering experiment.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
For AI-enhanced composites (digitalSourceType compositeWithTrainedAlgorithmicMedia,
identify ai_source_kind == "enhanced"; roadmap P1#8): regenerate ONLY the AI
region and preserve the real photo elsewhere, instead of regenerating the whole
frame.
- noai.tiling.feather_region_composite(base, regenerated, box, *, feather): pure,
model-free compositor that blends the regenerated AI box back over the original
with a feathered seam, leaving pixels OUTSIDE the box exactly equal to base.
Fully unit-tested (outside-box exactness, interior == regenerated, hard paste at
feather 0, monotonic seam ramp, dtype/grayscale/clamp/empty-box/shape-mismatch).
- WatermarkRemover.remove_watermark(region=, region_feather=) and the module-level
convenience function thread it through: the remover regenerates (or tiles) the
frame, then composites only the AI box back over the original input. The box is
caller-supplied -- a C2PA composite manifest carries no reliable machine-readable
region, so none is fabricated. The no-model lossless region path stays
region_eraser.erase.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Port the Gemini sparkle dark-pit guard (commit 41f6797) to the shared
TextMarkEngine reverse-alpha base (roadmap P0#8): on a dark or mid-tone
background the captured alpha can over-estimate this image's mark opacity, and
reverse-alpha leaves a darker-than-background glyph ghost instead of recovering
the true pixels. The sparkle-only fix left the text marks unhandled.
_reverse_alpha_oversubtracts predicts the reverse-alpha output PER PIXEL over the
glyph body from the INPUT ((obs - a*logo)/(1-a), the remover's own math); when
the predicted body lands more than _OVERSUB_DARK_MARGIN (25) gray levels below
the local background ring it abandons the reverse-alpha output for the footprint
and inpaints it from the original surroundings (_inpaint_footprint, wider dilate/
radius than the thin residual pass). Predicting per-pixel from the input (not the
produced output, which depends on which placement the remover picked) keeps a
cleanly captured full-strength mark byte-identical -- it predicts back to the
background everywhere, so the guard never trips on it (verified across all three
engines on white/mid/dark/midgray backgrounds).
Regression-guarded by tests/test_text_mark_oversubtraction.py: predicate True on
faint / False on clean, end-to-end no-dark-pit acceptance, clean-mark byte
identity, and textured-background footprint recovery.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Retained-corpus mining (2026-06-20) surfaced three provenance gaps; all are
oracle-free and regression-guarded.
- C2PA vendor coverage (roadmap): register Volcano Engine under its Chinese
legal entity 北京火山引擎科技有限公司 (the latin "volcengine" needle misses
those certs) -> normalizes to the same ByteDance platform; register ElevenLabs
("Eleven Labs Inc.", pure generative-AI) as a generator. Document the
deliberate exclusion of TikTok Inc. and PixelBin.io/"Fynd" (provenance/transform
signers, not generators) so they are not re-added.
- AI-generated vs AI-enhanced (roadmap): ProvenanceReport.ai_source_kind splits
the C2PA digital-source-type into "generated" (trainedAlgorithmicMedia) vs
"enhanced" (compositeWithTrainedAlgorithmicMedia) so a caller branches a
full-frame scrub from a region-targeted clean. Parsed once in
noai.c2pa._populate_registry_fields (PNG + any c2pa-python-readable container),
with a raw head-scan fallback in identify for the non-PNG raw-blob path. CLI
verdict reads "AI-generated (fully synthetic)" vs "AI-enhanced (real content
with an AI-composited region)"; surfaced in --json.
- Detect-vs-remove threshold desync (P0#7): identify's sparkle threshold and the
removal arbitration gate were two independent 0.5 constants. Unify them into the
single GEMINI_SPARKLE_TRUST_CONF (identify imports it) so they can never drift.
Lowering the gate to recover faint sub-0.5 sparkles was evaluated and REJECTED:
a real Doubao text mark scores ~0.40-0.42 as a gemini match with a higher
core-ring brightness margin than a genuine faint sparkle, so neither confidence
nor the brightness gate separates them in [0.35, 0.5) -- lowering would trade a
rare miss for false-positive removals on clean images. Regression-guarded by
TestSparkleDetectRemoveAlignment (real demo sparkle at borderline opacities;
identify and best_auto_mark must agree on either side of the line).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Measured (openai_1, 0.10, seeds 0-4): seed barely moves whole-image fidelity
(img LPIPS 0.062-0.065, SSIM/PSNR flat) but shifts text legibility (OCR CER
0.241-0.290, ~17% spread) -- it changes which details regenerate, not the level.
So per-image best-of-N-seed is a weak text-only lever (pin a seed in prod; reserve
best-of-N for text-heavy premium). Also retitle the qwen section "certified floors"
and drop the now-stale "uncertified / run seed-repeat / floor 0.30" tails.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Oracle seed-repeat + floor refinement (2026-06-20, data/qwen_in):
- OpenAI floor 0.10 is SEED-ROBUST: 0.05 and 0.075 still detected; 0.10 clean on
seeds 0-4 (5/5) -> a random seed is safe.
- Gemini floor lowered 0.30 -> 0.25 (0.20 still detected, 0.25 clean on both
images). Single-seed (seed 0): the Gemini oracle rate-limits volume seed-repeat,
so pin a seed in prod rather than relying on seed-robustness there.
Re-measured fidelity at the certified floors (controlnet 0.15 vs Qwen 0.25 for
Gemini): faces still favor controlnet (ArcFace 0.546 vs 0.382, lapvar 0.62 vs
0.40); the short-CJK text case is now a TIE (gemini_1 0.037 vs 0.037 -- the earlier
Qwen 0.000 was at 0.30, not the floor). Qwen's text win holds on substantial
Latin/mixed text (OpenAI 0.385 vs 0.241 / 0.341 vs 0.290). Update watermark_profiles
comment, CLAUDE.md, module-internals, known-limitations.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The face fidelity numbers cited an equal-strength compare (both 0.15), but Qwen at
0.15 does NOT clear Gemini SynthID -- so that output is un-scrubbed and the compare
is invalid. Per the methodology rule (compare fidelity only between outputs where
SynthID is removed in BOTH), restate faces at each pipeline's scrub floor
(controlnet 0.15 / Qwen 0.30): ArcFace identity 0.546 vs 0.331, lapvar 0.62 vs 0.40,
face LPIPS 0.09 vs 0.19 -- controlnet still wins faces, conclusion unchanged. Drop
the "equal strength" framing in CLAUDE.md / module-internals / known-limitations.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
data/qwen_in/ground_truth.json is transcribed by vision (PaddleOCR mangled the
stylized Cyrillic), so the text metric scores variants against an accurate
reference instead of noisy OCR-vs-OCR. Re-measured text CER (controlnet vs qwen)
with this ground truth confirms qwen wins text across EN/RU/ZH: openai_1 0.385 vs
0.241, openai_2 0.341 vs 0.290, gemini_1 (ZH) 0.037 vs 0.000 (perfect Chinese even
at the higher 0.30 strength). Faces still favor controlnet. Refresh the numbers in
docs/known-limitations.md to this cleaner methodology.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- data/qwen_in/: a stable, committed set of 4 AI-generated images (OpenAI +
Google, carrying SynthID/C2PA -- same class as data/samples fixtures) used to
compare the controlnet/sdxl/qwen pipelines for fidelity. Two text-multi-script
(incl. RU/CJK), one EN poster, one face grid. README documents the set + the
ground-truth workflow. data/ is sdist-excluded so the wheel is unaffected.
- scripts/fidelity_metrics.py: switch text OCR from EasyOCR to PaddleOCR
(PP-OCRv6, higher accuracy esp. CJK, single multilingual stack); split into
`ocr` (seed a {basename: text} ground truth) and `compare` (--ground-truth for
a clean CER vs the hand-verified reference instead of noisy OCR-vs-OCR). Spatial
IoU-NMS keeps the best-scoring read per line so wrong-script models don't inject
garbage over Cyrillic/CJK.
- Oracle methodology: validate the OpenAI arm FIRST (openai.com/verify is more
accessible and the strongest Playwright/Chrome-MCP automation candidate; the
Gemini app is more manual). Recorded in CLAUDE.md + docs/synthid.md.
Ground-truth JSON (data/qwen_in/ground_truth.json) lands in a follow-up once
hand-verified.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add scripts/fidelity_metrics.py: an objective eval harness comparing
watermark-removal outputs against the original (reference) across four groups
-- OCR character error rate (EasyOCR), ArcFace identity cosine (insightface),
face texture (LPIPS + Laplacian-variance ratio), and whole-image LPIPS/SSIM/
PSNR. PEP 723 inline deps so it stays out of the package / uv.lock; metrics
self-gate (faces only where faces, text only where text).
The metrics overturned an eyeball conclusion: at EQUAL strength Qwen beats
controlnet on TEXT (OpenAI typography 0.10: OCR CER 0.25 vs 0.37) but controlnet
beats Qwen on FACES (gemini_3, 18 faces, 0.15 each: Laplacian-variance retention
0.62 vs 0.41, face LPIPS 0.09 vs 0.13 -- Qwen smooths faces MORE; ArcFace
identity ~tied). So Qwen is the better TEXT-preserving remover, not a universal
fidelity win. Correct the earlier "qwen keeps faces faithful where controlnet
plasticizes" claim in CLAUDE.md, module-internals.md, known-limitations.md, README.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A third diffusion pipeline alongside sdxl/controlnet: Qwen-Image (20B MMDiT,
Apache-2.0 code AND weights) img2img. The scrub still comes from the img2img
strength; Qwen preserves text (incl. CJK) and structure markedly better than
SDXL at the scrub floor, so it over-regenerates real photos far less (directly
targets the controlnet over-regeneration that degrades real uploads).
- watermark_profiles: QWEN_MODEL_ID, normalize_profile accepts "qwen".
- WatermarkRemover: _load_qwen_pipeline (bf16, loads Qwen base unless --model
overridden, clear ImportError if diffusers lacks the class), _run_qwen (no
MPS fallback -- 20B is CUDA/cloud-class), dispatch in _generate_one/preload,
pure _build_qwen_kwargs (true_cfg_scale, not guidance_scale).
- Shared _base_load_kwargs() across all three loaders (dtype + token).
- CLI --pipeline gains "qwen"; invisible_engine threads it through.
- scripts/qwen_scrub_prototype.py: standalone PEP 723 GPU experiment.
Prototype oracle floors (Modal A100-80GB, single seed, controls SynthID-positive,
PENDING seed-repeat cert): OpenAI clears at strength ~0.10, Gemini at ~0.30 (0.20
still detected), with CJK text + faces faithful where controlnet plasticizes. The
Gemini floor is higher than the shared default ladder, so pass an explicit
--strength for Gemini on this pipeline until a Qwen-specific ladder is certified.
The model-running path is CUDA-only (untestable locally); unit tests cover the
pure call-shape (_build_qwen_kwargs) and profile normalization without torch.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add a lossless alternative to the --max-resolution downscale for large
images that OOM on MPS/GPU: regenerate in overlapping, feather-blended
tiles at native resolution.
- noai/tiling.py: pure plan_tiles (uniform tiles, last flush to edge) +
feather_weights (strictly-positive separable taper -> partition-of-unity
blend) + run_tiled (per-tile generate callable, decoupled from the
pipeline). Unit-tested without the model.
- WatermarkRemover.remove_watermark: refactor _generate into _generate_one
+ a tiled branch that engages only when --tile is set and the long side
exceeds tile_size (ControlNet canny is rebuilt per tile).
- Thread tile/tile_size/tile_overlap through InvisibleEngine and the
invisible/all/batch CLI commands via a shared _tile_options decorator.
Verified end-to-end on the real SDXL pipeline (forced 2x2 tiling on a
1024px sample, MPS): non-degenerate output, no gross seam at tile borders.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Closes a documented coverage gap (P2#9): an AI Software/Make/Artist/ImageDescription
token in an EXIF item (its TIFF bytes live in mdat/idat) survived remove_ai_metadata
because the top-level box stripper and (absent pillow-heif) the PIL EXIF reader can't
reach it. New isobmff.blank_ai_exif_tokens finds EXIF TIFF blocks by their II/MM
byte-order header, validates each with piexif (a coincidental II/MM run in pixels
won't parse as a TIFF IFD, so it's ignored), and overwrites any AI_GENERATOR_TOKENS-
bearing value with same-length spaces -- so box sizes and iloc offsets stay valid and
the coded image is untouched (mirrors blank_ai_xmp_packets; no iinf/iloc surgery, no
exiftool dep). Camera/editor EXIF without an AI token is preserved. Wired into
remove_ai_metadata's ISOBMFF path. Covers the realistic AI-generator-token case; xAI-
signature-in-meta-box-EXIF (Grok is JPEG-only) stays out.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Deep-research (2026-06-19, adversarially verified) confirms the open imwatermark
dwtDct mark is fragile by scheme, not by our usage: maintainers admit no 100%
clean-decode guarantee; measured ~0.79 bit accuracy clean (~38/48, below our 44
gate). Root causes (code-verified + locally reproduced): per-block max-coefficient
bit read (content flips bits) and YUV chroma 8-bit clamping on bright pixels (the
bright-flat / all-ones failure). No maintained fork or detector does this scheme
reliably (WAVES relegates it to an appendix; learned schemes are a different class;
dwtDctSvd cannot decode SDXL's dwtDct). Conclusion: keep it positive-only, rely on
C2PA. Sources: imwatermark READMEs, arXiv:2406.08337 (WMAdapter), arXiv:2401.08573
(WAVES), diffusers SDXL watermark.py.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Final characterization after a positive-control sweep. The imwatermark dwtDct
round-trip fails (28-39/48, below the 44 gate) not on "high texture" as a prior
note claimed, but on a broad carrier class: the FLUX fox, doubao, a minimalist-FLAT
FLUX generation, AND a clean synthetic bright-flat fill with NO watermark all fail
identically. The degenerate all-ones decode is therefore a CARRIER ARTIFACT, not a
watermark (the no-watermark synthetic image reproduces it; a double-embed test shows
no interference). detect_invisible_watermark is positive-only: trust a hit, treat a
None as inconclusive unless a same-carrier positive control first recovers >=44.
Consequence: whether BFL hosted FLUX embeds the open DWT-DCT is unresolvable with
this detector on the available carriers (textured AND flat FLUX both fail the
control). C2PA stays the reliable FLUX signal. Low priority to chase further.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Victor Kuznetsov