Sora characters gone: reuse a subject with reference images

Sume has no saved character object for video. To keep one subject across shots, send the same reference images in input_references on every request to a model that accepts them: the Seedance 2.x models, wan-3.0, minimax-h3 or minimax-h3-max. Nothing is stored between jobs, so the images travel with each request.

OpenAI's guide described Characters as reusable non-human subjects across generations, with 2 to 4 second source clips working best and human likeness blocked by default, in an API that was shut down on 2026-09-24 (read 2026-10-01).

Which Sume models take reference images?

The video docs say audio and video references are honored by the Seedance 2.x models, Wan 3.0, MiniMax H3 and MiniMax H3 Max, and that input_references triggers reference-to-video. The catalog lists per-model image limits:

How do I send a subject on every request?

Host the image at a public HTTPS URL and repeat the entry. The prompt should name what the reference shows, since the model uses the images as guidance rather than exact frames.

curl -X POST https://api.sume.com/v1/videos \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: mascot-shot-02" \
  -d '{
    "model": "wan-3.0",
    "prompt": "The red fox mascot from the reference walks into a bakery",
    "resolution": "720p",
    "duration": 6,
    "input_references": [
      {
        "type": "image_url",
        "image_url": { "url": "https://example.com/fox-mascot.png" }
      }
    ]
  }'

Does this work for people?

Treat it as a stylized-subject workflow. Seedance rejects real-human face references; see what to do when a reference face is rejected. Cross-shot consistency is the model's best effort, not a guarantee, so review each clip. More patterns are in consistent character across AI video shots, and the field details are in the video docs.