C2PA label for a partly AI-edited video: compositedWith...

If an existing clip is changed by a model instead of generated from scratch, the C2PA 2.3 guidance points to a different IPTC source type: http://cv.iptc.org/newscodes/digitalsourcetype/compositedWithTrainedAlgorithmicMedia. The guidance gives inpainting as its example of an AI-assisted edit.

The split matters for labeling. A text-to-video clip and an edited camera clip are not the same claim, and a Content Credential can say which one applies. This post covers the C2PA side and which Sume jobs fall on each side of it, using the Sume docs only. The docs do not say Sume writes a manifest.

Which Sume jobs are edits

Per the Sume video docs, the Gemini Omni Flash model exposes a video edit mode through the Video Router video_url field. Face swap is a separate Beta endpoint that takes avatar_handle, video_url and quality. Both start from an existing video, so a provenance record should treat the output as an edit of that input.

What to log per edit

The implementation guide from C2PA lists c2pa.created, c2pa.opened and c2pa.edited as the actions to use. An edit chain therefore opens the source, then records the AI step. The source clip is an ingredient of the output.

• The source artifact URL that went in as video_url.
• The Sume job id and artifact URL that came out.
• The model id or endpoint used, for example the face swap endpoint.
• The date and time of the submit.

Cost context

Edits on video models follow the same rule as generation: provider list times 1.25, reserved on submit. A plain trim is different, a flat $0.02 per job, and it uses no model. Keeping the two apart in the log keeps the label honest and the invoice readable. See video trim.