Evaluate selective text restoration

docs/known-limitations.md

@@ -61,8 +61,25 @@ change even when the watermark is successfully disrupted.
 Both are CUDA only and differ only in the global regeneration model: each
 conditions that stage on a canny edge map, which preserves structure but not
 identity or exact texture, and each then runs the same face stage.
-`qwen-zimage` is the higher fidelity of the two. Both are large, slow, and may
-still alter small text or difficult faces.
+Existing face evaluations favor `qwen-zimage`, but there is no blanket fidelity
+ordering across content types. A fixed-seed, three-scene text comparison at the
+profile defaults found no stable winner: SDXL won one poster, Qwen won one, and
+the Chinese sign tied. Both are large, slow, and may still alter small text or
+difficult faces. The measurements and their OCR and oracle caveats are tracked
+in [`data/evaluations/fidelity/`](../data/evaluations/fidelity/README.md).
+A global Z-Image Turbo prototype preserved text substantially better at low
+strength, but it has no useful cross-provider operating point and is not a
+supported profile. The evaluated text restorers also remain research-only:
+fresh-font and silhouette variants visibly changed typography, while the
+higher-fidelity `vae-glyphs` route still requires verified strings, line
+geometry, a separately generated donor, and an independently clean global
+anchor. Automatic OCR and line-box proposals are not reliable enough to remove
+those requirements, and the exact oracle results do not establish a general
+mask, seed, 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
+[`text-protection-research.md`](text-protection-research.md) and the
+[`fidelity` evaluation record](../data/evaluations/fidelity/README.md).
 
 ### Removal cannot be verified locally for proprietary SynthID
 
@@ -175,6 +192,14 @@ certified at a fixed seed. The live resolver is
 | `qwen-zimage` | CUDA only, large model stack, and limited broad certification across seeds and content. |
 | `sdxl-zimage` | CUDA only. Its strength ladder is flat per vendor, not a resolution curve, because flat values are what was measured. |
 
+The evaluated text-restoration prototypes are not optional production stages.
+OCR plus LaMa recovered literal poster text but changed fonts and worsened whole-image
+fidelity. Restricting it to OCR-mismatched lines improved the tradeoff but still
+left a local shadow on one poster. The published AnyText2 SD1.5 checkpoint
+substituted Chinese characters and increased CER on the sign fixture. AnyText2XL
+is not published, and the released wrapper truncates individual text lines after
+20 characters.
+
 The `controlnet`, `sdxl`, `qwen` and `default` profiles were removed, not aliased
 onward: a retired name is rejected at parse time rather than routed into a profile
 the caller never chose. There is no `--model`, `--steps`, `--guidance-scale`,

docs/module-internals.md

@@ -949,6 +949,34 @@ orchestration, YuNet integration, SAM selection, masks, sizing helpers, and pixe
 compositing are implemented for this runtime. Changing a calibrated model input
 requires the same provider-oracle and identity evaluation as a model change.
 
+A matched stage-isolation check on the 18-face Gemini portrait grid confirms the
+division of responsibility. The visible-cleaned, metadata-stripped control and the
+Z-Image face-only output were both SynthID-positive; Qwen global-only and the full
+Qwen-then-Z-Image output were both clean. The face stage raised identity cosine from
+0.589 to 0.852 and reduced face LPIPS from 0.217 to 0.050 without reintroducing a
+detectable whole-image signal. Thus Z-Image is a masked fidelity repair stage here,
+not the watermark-removal stage. Exact hashes, metrics, strengths, and the one-fixture,
+one-seed caveat are recorded in
+[`data/evaluations/fidelity/face-stage-isolation-2026-08-13.csv`](../data/evaluations/fidelity/face-stage-isolation-2026-08-13.csv).
+
+The public Synthid-Bypass v2 graph was subsequently audited at upstream commit
+`3007d035`. Its saved-output path confirms the same division: Qwen-Image-2512
+Lightning plus Canny is global, and Z-Image Turbo exists only inside the masked
+face detailer. The connected face path is YOLOv8-face plus SAM; the MediaPipe nodes
+described by the upstream README and the 1.2-megapixel normalization node do not
+reach `SaveImage`. Upstream also applies its adaptive face strength directly,
+whereas this implementation multiplies it by `FACE_DENOISE_SCALE = 0.5`.
+
+A close reproduction on the same portrait fixture kept the control positive and
+made both Qwen global-only and full Qwen-then-Z-Image outputs clean. Applying the
+upstream-strength face pass raised identity from 0.589 to 0.783 and reduced face
+LPIPS from 0.217 to 0.083, but remained worse than this profile's 0.852 and 0.050.
+The published upstream pair 12 was also independently checked positive before and
+clean after, with 0.975 identity. Exact workflow provenance, hashes, metrics,
+oracle outcomes, and the DiffSynth/GGUF, scheduler, detector, and seed caveats are
+recorded in
+[`data/evaluations/fidelity/upstream-v2-reproduction-2026-08-13.csv`](../data/evaluations/fidelity/upstream-v2-reproduction-2026-08-13.csv).
+
 ### SDXL plus Z-Image
 
 [`_internal/sdxl_zimage_pipeline.py`](../src/remove_ai_watermarks/_internal/sdxl_zimage_pipeline.py)

docs/text-protection-research.md

@@ -143,3 +143,203 @@ Deep-research workflow run `wf_118b9a03-3eb` (2026-05-29). Findings adversariall
 (2/3 refutes required to kill a claim). This note records research only; no code change is
 implied until a prototype validates fidelity and the SynthID-scrub guarantee on the
 restored output.
+
+## 2026-08-13 OCR plus LaMa prototype
+
+A lightweight version of the recommended two-stage architecture was tested on
+the three tracked text fixtures. It used the already-clean Qwen result as the
+base, recognized the source text, removed source and Qwen glyph contours with
+LaMa, and rendered the recognized strings as fresh pixels. No source pixels
+were pasted back.
+
+The result validates only part of the hypothesis. Character-weighted CER on the
+two multilingual posters fell from 0.338/0.305 to 0.007/0.007, and OpenAI Verify
+remained clean for both matched Qwen and restored pairs. However, replacement
+fonts changed the design, whole-image LPIPS worsened by 0.067 on each poster,
+and the light poster retained a shadow in one small English line. On the Chinese
+sign, punctuation errors moved CER from 0.074 to 0.111. The Google verifier run
+was inconclusive because the available account detected both the restored image
+and the byte-identical Qwen control that a previous account had called clean.
+
+The experiment supports a proper glyph-conditioned restorer, not shipping the
+plain OCR/system-font compositor. Exact hashes, metrics, masks, and provider
+verdicts are tracked in
+[`data/evaluations/fidelity/text-restoration-2026-08-13.csv`](../data/evaluations/fidelity/text-restoration-2026-08-13.csv).
+
+## 2026-08-13 AnyText2 follow-up
+
+The official AnyText2 SD1.5 checkpoint was then tested as the glyph-conditioned
+restorer. Its ModelScope entry and code are Apache 2.0, and the downloaded
+checkpoint matched the published SHA-256. An official blackboard edit completed
+successfully before the fixture run, establishing that the runtime reproduced
+the model rather than silently exercising a fallback.
+
+On the Chinese sign, a three-line local edit over the exact Qwen output scored
+CER 0.185 under the standard detector. Font mimic from the source glyph masks
+scored 0.222. Whole-image LPIPS changed from 0.289 to 0.338 and 0.345 respectively.
+A padded crop-recognition check showed why the comparison matters: Qwen contained
+the full correct text and scored 0.000, while default AnyText2 still scored 0.074
+with two substituted characters and mimic remained at 0.222. The provider oracle
+was deliberately not called because these variants had already failed the
+content-fidelity gate.
+
+AnyText2 is therefore not the missing production restorer in its published form.
+The negative result is narrower than the model family: only the SD1.5 checkpoint
+is public, while AnyText2XL remains listed as future work. The current wrapper
+also truncates each quoted text line to 20 characters, which prevents a direct
+test of several tracked English and Russian poster lines. Exact model provenance,
+output hashes, and metrics are recorded in
+[`data/evaluations/fidelity/anytext2-restoration-2026-08-13.csv`](../data/evaluations/fidelity/anytext2-restoration-2026-08-13.csv).
+
+The same cross-check exposes a measurement bug in the earlier lightweight
+restoration result. Paddle detection made tight boxes that omitted the final
+Chinese full stop from two lines, producing Qwen CER 0.074 although the pixels
+and padded recognition were correct. Adding 16-32 pixels of horizontal context
+restored every punctuation mark. A deterministic rerender using the full OCR
+strings and the closest of nine mask-scored CJK fonts also reached detector CER
+0.000, but its heavier Hiragino Sans GB W6 glyphs raised LPIPS from 0.289 to 0.327.
+The actionable design is selective restoration: compare padded source and output
+recognition first, and preserve the Qwen output when they already match.
+
+That policy was tested on the two multilingual posters. A manual prototype kept
+the lines whose padded source and Qwen recognition agreed and rerendered only the
+changed lines. The edited mask fell from 15.4%/17.0% in the full compositor to
+5.5%/5.5%. On a single Paddle `en+ru+ch` route, CER changed from 0.378/0.413 for
+Qwen to 0.101/0.112, while LPIPS was only 0.120/0.103 rather than the full
+compositor's 0.174/0.162. The different OCR route is intentional and all three
+variants were remeasured through it; these CER values are not directly
+comparable with the earlier Vision/Paddle hybrid results.
+
+Both selective outputs returned `No OpenAI signals detected` in OpenAI Verify,
+and the original light poster returned `Generated with OpenAI tools` immediately
+afterward as the positive control. The dark poster was visually clean. The light
+poster still showed one local old-text shadow, so mask construction remains the
+blocking defect. Exact hashes, metrics, mask fractions, and oracle controls are
+in
+[`data/evaluations/fidelity/selective-text-restoration-2026-08-13.csv`](../data/evaluations/fidelity/selective-text-restoration-2026-08-13.csv).
+
+Uniformly expanding every selected glyph mask by two pixels removed that light
+poster shadow and moved CER from 0.101 to 0.081 without a meaningful LPIPS cost.
+It also expanded the edited area from 5.5% to 7.0%, and the identical rule made
+the dark poster worse at CER 0.147 rather than 0.112. Those follow-up bytes have
+not been oracle-checked. The next prototype should detect residual source glyphs
+per line and expand only the failing component, rather than introducing another
+poster-wide tuning constant.
+
+The manual selection was then replaced with per-line padded recognition. A line
+is left untouched only when the source recognizer is sufficiently consistent
+with the verified line and normalized source/candidate recognition agree
+exactly. The automatic rule reproduced the manual selection on the dark poster
+and correctly kept one additional large Russian line on the light poster. It
+reduced that poster's effective edited area from 5.5% to 4.1%.
+
+Mask cleanup became a measured second pass rather than a global dilation. After
+the first LaMa erase, the prototype finds contrast that remains specifically at
+the original glyph positions, dilates only that residual, and erases it before
+fresh text is drawn. This removed the visible double glyphs from both posters.
+The automatic outputs scored LPIPS 0.113/0.104 and detector CER 0.123/0.119, with
+effective edited fractions of 4.1%/5.7%. The light aggregate CER is pessimistic:
+the page detector fragments its untouched Cyrillic line into Latin-like pieces,
+while the padded Russian recognizer returns the exact expected text.
+
+OpenAI Verify returned `No OpenAI signals detected` for both new hashes and then
+`Generated with OpenAI tools` for the original light poster in the same browser
+sequence. Reproducing the prototype from the tracked script exposed harmless
+ONNX rounding on the light image: 386 pixels differed by at most one channel
+value from the initially checked hash. The exact tracked output hash
+`900def5a...` was therefore submitted separately, returned clean, and was
+followed immediately by another positive source result. The dark tracked output
+was byte-identical to the initially checked hash.
+
+The Chinese sign provides a no-edit control for the selector. Detector boxes for
+its three large lines overlap vertically, so Latin-style vertical padding
+incorrectly mixed neighboring lines. Horizontal-only CJK padding captured the
+terminal punctuation without mixing neighbors and made source and candidate
+recognition agree exactly on all three lines.
+The script then produced a zero mask and copied the Qwen candidate byte for byte.
+This prevents the heavier-font regression seen in the earlier CJK compositor,
+but does not resolve the unavailable Google-negative oracle control.
+
+The result clears the measured OpenAI gate without a manual line selection, but
+is not a production design yet. It still assumes verified source strings and
+line boxes, relies on macOS system fonts, and needs evaluation on uncurated
+layouts, rotated text, false OCR matches, automatic line-box discovery, and a
+reproducible Google-negative control.
+
+The later source-silhouette experiment removed the font lookup but did not meet
+the actual visual requirement. Native-size review found changed stroke weight,
+color variation, edge antialiasing, and decorative details even where crop OCR
+was exact. A narrower `vae-glyphs` experiment uses the silhouette only as a
+support mask: exact pixels come from a Qwen-VAE reconstruction, while a fresh
+flat-color silhouette supplies only the outer edge beneath that core. Across 48
+typography cases (548 annotated lines), this improved glyph-edge MAE in 48/48
+and text-box SSIM in 47/48. The median values moved from 37.59 to 32.42 and from
+0.854 to 0.914 respectively. One separate untracked core-only output with a
+1.43% donor area returned `No OpenAI signals detected`; the tracked script then
+reproduced those exact bytes. A 0.5-pixel feathered sibling with 2.75% nonzero
+donor area improved mean text-box SSIM from 0.957 to 0.960, with a 0.918 minimum
+across 15 verified lines, while full-image RGB SSIM reached 0.864. Crop OCR
+recovered from 7/15 exact lines on the raw pass to 14/15, equal to the source's
+own recognizer floor. Whole-image LPIPS was 0.082, but only 0.108% of pixels
+remained exactly equal to the source and the detected face's Laplacian-variance
+ratio was 0.670, confirming residual global smoothing. A same-session
+OpenAI Verify sequence detected the exact source control in 1/1 check, returned
+clean for the raw global pass and clean-fidelity anchor in 1/1 each, and returned
+clean for the feathered output in 3/3 repeats. The tracked script reproduced
+that feathered file byte for byte. This proves two materially better exact-output
+Pareto points, not a general operating range. The 48 matrix outputs and other
+mask sizes remain oracle-unverified.
+
+The first Google oracle sample did not transfer. For the synthetic CJK sign,
+two separate work accounts detected SynthID in both the resaved positive source
+control and the exact Qwen-VAE donor candidate. The candidate's text-box SSIM
+improved from 0.791 to 0.835 and its glyph-edge MAE from 35.48 to 22.49, but its
+3.67% donor layer did not create a clean operating point. The Qwen silhouette
+base was already detected, as was the earlier original-sign Qwen baseline in the
+same account. The measured cause is therefore insufficient global Google
+regeneration, not a demonstrated watermark regression from the text donor.
+Google needs a stronger oracle-clean global anchor before the narrow donor can
+be evaluated there.
+
+A targeted follow-up supplied that missing anchor for one CJK case. An explicit
+Qwen global pass at strength 0.30 returned no reliable SynthID signal in Gemini,
+while the matched source control was detected. Applying the 0.5-pixel
+`vae-glyphs` layer to that clean anchor changed 1.17% of the image and raised
+mean text-box SSIM from 0.765 to 0.964 (minimum 0.963), with mean text-box MAE
+falling from 20.74 to 5.38. The post-donor output then returned clean in 3/3
+Gemini checks across two work accounts. A separate 18-face fixture with four
+tiny UI-text lines also had a clean strength-0.30 anchor; its donor layer changed
+0.70% of the image, raised mean text-box SSIM from 0.617 to 0.976 (minimum
+0.972), and returned clean in 3/3 checks across the same two accounts. One
+quota-exceeded response is excluded from both pass rates. These are two exact
+oracle-certified outputs, not a general operating range: other layouts, masks,
+strengths, seeds, and donor areas remain unverified.
+
+An automatic-box follow-up merged Paddle word detections by vertical overlap.
+It found exactly 20/20 poster lines and 3/3 sign lines, with mean IoU 0.857,
+0.847, and 1.000 against the verified boxes. That structural match was not
+sufficient: reusing the annotation crop padding changed recognition decisions,
+reduced preserved dark-poster lines from 7 to 4, and expanded its edited
+fraction from 5.7% to 11.2%.
+
+A recognition-only sweep isolated the crop sensitivity. Limiting vertical
+padding around detector boxes to 8-12% of line height reproduced the verified-
+box decision vectors on both posters; 10% was used for a full follow-up. It kept
+8/20 and 7/20 lines with edited fractions of 4.1% and 5.7%, and visual inspection
+found no double glyphs. Whole-image LPIPS remained 0.113/0.104, but detector CER
+was 0.127/0.154 instead of 0.123/0.119. The dark regression failed the fidelity
+gate, so these hashes were not submitted to the provider oracle. The opt-in flag
+remains only to reproduce the negative result. Count, IoU, and matching decision
+vectors are therefore insufficient gates for automatic boxes; the next design
+needs box rectification or recognition stability under crop jitter before it
+can replace verified geometry.
+
+Removing verified strings was tested separately with an annotation-seed dry
+run. It detects boxes, chooses `en`, `ru`, or `ch` from Unicode script, and
+accepts a draft only when three crop paddings normalize identically and every
+confidence is at least 0.85. One execution proposed 20 and 18 poster lines, but
+exact-text precision against the held-out annotations was only 90.0% and 94.4%.
+The stable errors were punctuation: one lost English comma and an ideographic
+comma consistently replaced by ASCII. Thus crop agreement and confidence do
+not verify literal text. The helper is retained only to seed annotations; every
+proposal needs manual review before it can enter the restoration compositor.