frame_images plus input_references in one request: which wins?

If a POST /v1/videos request carries both frame_images and input_references, frame_images takes precedence and the job is treated as image-to-video. The references are then not used as the generation mode, so send only one of the two.

What is the difference between the two fields?

frame_images gives first or last frame images for image-to-video; each entry needs a frame_type of first_frame or last_frame. input_references gives style or content references for reference-to-video, and the model uses them as visual guidance rather than exact frames.

Which mode does each field trigger?

The docs tie each field to one generation mode.

How do I send a first frame?

Use a public HTTPS image URL and mark it first_frame. The example follows the docs shape.

curl -X POST https://api.sume.com/v1/videos \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: frame-001" \
  -d '{"model":"seedance-2","prompt":"A character walking through a forest","frame_images":[{"type":"image_url","image_url":{"url":"https://example.com/first-frame.png"},"frame_type":"first_frame"}],"resolution":"1080p"}'

Does every model accept every reference type?

No. Only models whose supported_input_references lists a type accept it, and supported_frame_images says which frame_type values a model takes. Check GET /v1/videos/models first; details are in the video docs.

What happens after the job is accepted?

The submit response is 202 Accepted with an id, polling_url, status of pending and the model. Video generation is asynchronous: submit, receive a job id and polling URL, poll until completed, then download from the content URL. The docs suggest a polling interval of about 30 seconds and say generation typically takes 30 seconds to several minutes depending on the model and parameters.

If a request fails, check that any image is reachable over public HTTPS and in a supported format, as the troubleshooting notes advise.