Skip to content

v0.6.11: detect China TC260 AIGC PNG chunk + HuggingFace hf-job-id

Choose a tag to compare

@wiltodelta wiltodelta released this 28 May 19:49
· 145 commits to main since this release
v0.6.11
f326bab

v0.6.11: detect China TC260 AIGC PNG chunk and HuggingFace hf-job-id

New provenance signals

Auditing a 646-image corpus surfaced two markers the library previously reported
as Unknown. identify now catches both:

  • China TC260 AIGC label as a PNG tEXt chunk. Doubao (ByteDance) and other
    China-served generators also write the mandatory TC260 label as a raw-JSON
    AIGC PNG chunk, with no namespaced XMP marker. aigc_label now reads both
    serializations (the XMP <TC260:AIGC> block and the PNG chunk) via a shared
    parser, gating the PNG path on a known TC260 field so a generic AIGC key
    cannot false-positive. 28 corpus files now resolve as AI (high confidence).
  • HuggingFace-hosted job marker (hf-job-id). HuggingFace Jobs / Spaces
    stamp generated PNGs with an hf-job-id chunk. It marks the hosting job, not
    a model (commonly diffusion output), so it lifts an Unknown verdict to a
    tentative AI at medium confidence, parallel to the visible sparkle and
    never overriding a hard metadata signal. 3 corpus files.

Both markers are stripped on removal (the PNG save whitelist keeps only standard
metadata keys) and surfaced by has_ai_metadata / get_ai_metadata.