Remove AI Label | Public file tests | 2026-11-07

METADATA
Each original is a frozen input. Download it, run Privacy clean in the local
tool, and compare your output's SHA-256 with wholeFileSha256.after in the JSON
report (use Keep EXIF credits for camera-credits-report.json). On macOS, use shasum -a 256 FILE. On Windows PowerShell, use
Get-FileHash FILE -Algorithm SHA256. The two whole-file hashes can differ
even when the embedded media hashes and decoded pixels match.

Eight Privacy clean outputs and one Keep EXIF credits output were also
decoded independently. The credits example uses the same camera JPEG. The image checks used
Sharp/libvips, applied EXIF orientation and compared full RGBA pixel buffers
and ICC profile bytes. The video checks used FFmpeg to decode both streams
and compare the framemd5 rows, including timestamps. Each success report
contains that check. ExifTool independently read selected GPS and author
fields in the JPEG and WebP tests. The damaged PNG report contains the error
returned by the production cleaning engine.

In a checkout with this evidence set and project dependencies, run:
  pnpm exec tsx scripts/build-evidence.ts 2026-11-07
Before running, save the SDXL Turbo source PNG linked below to
artifacts/evidence/comfyui-original.png. The script reads this external
image and verifies its checksum. It keeps the resulting PNG in artifacts.
This rebuilds output copies, cropped repair previews, and reports from the
shipped original files. It runs offline. Output encoders can vary by version.
The C2PA SDK report records a separate verification of the frozen JPG pair.

INPUT SOURCES
Camera JPEG, transparency PNG, WebP, MP4, MOV and damaged PNG: generated test
patterns. The synthetic inputs and test tags are defined in
tests/metadata-engine.test.ts. The UUID and credential-like PNG chunks in
these synthetic files contain test strings rather than signed credentials.

Lake demo PNG: the site-owned image in public/images/sample-with-metadata.png,
with two illustrative text fields added by scripts/prepare-assets.ts.

ComfyUI workflow PNG: the upstream SDXL Turbo example published by
comfyanonymous, retrieved unchanged from:
https://raw.githubusercontent.com/comfyanonymous/ComfyUI_examples/f9431bb000ce792094ff345446e22cac1ea6cef3/sdturbo/sdxlturbo_example.png
The upstream workflow license is included alongside this file as
comfyui-license.txt. The checksummed original contains both prompt and workflow.

Signed JPEG: the generated camera chart with orientation and ICC profile,
signed locally with c2pa-python 0.38.2 and the SDK's public es256 test
certificate. Its claim records c2pa.created with digitalCreation as the
source type. No time-stamp service was requested. The SDK independently read
the original's active manifest, verified its signature and data binding,
reported the signing credential untrusted, and found no manifest in the
cleaned copy. See c2pa-reader.json. The browser tool only detects containers.
Public test certificate source:
https://github.com/contentauth/c2pa-python/tree/3c31c55bad632a59bb8c88ce4a54215a450eae82/tests/fixtures

REPAIR
Input: the site's generated 768 x 512 lake sample with a SAMPLE overlay.
The mask selects x=510, y=435, width=140, height=45. Crop previews preserve
the same 180 x 85 region in every file. The free output was regenerated using
the site's inpaintImage routine. The paid output is the original saved
composite from the site's November 7 LaMa acceptance run via Replicate,
using allenhooo/lama version:
cdac78a1bec5b23c07fd29692fb70baa513ea403a39e643c48ec5edadb15fe72

To compare the repairs independently, decode each full-size PNG as RGBA.
Compare the input pixels with each repaired output wherever the mask is
black. Every RGBA value should match there. Compare the white region against
the unmarked lake to see the changes in the reconstructed texture.