Make a shop mascot copy a dance: Genjutsu source video and images

To make a shop mascot copy a dance, use higgsfield-genjutsu on the Video Router: send your dance clip as video_url and 1 to 8 pictures of the mascot as reference_image_urls. It accepts 4 to 30 seconds at 480p or 720p. The model is Motion Transfer, so a request without a source video is not valid, and the docs list no aspect_ratio or generate_audio for it.

It is only listed when its provider is configured, so check GET /v1/video-router/models first. Source: Video Router, read 2026-10-04.

Which model for which job

Genjutsu is one of the two models on the Router that start from a source video. The other, h3-max-recast, swaps people for photos of people.

Prepare the source

Choose a dance clip with one dancer, a steady camera and the full body in frame. Probe it with POST /v1/video-inspect and frames: false so you know its length before you set duration in the 4 to 30 window.

Request

Show the mascot from the front and one side, on plain backgrounds. Up to 8 images are allowed, and more angles help only if they are consistent.

curl -X POST https://api.sume.com/v1/video-router/generate \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: mascot-dance-001" \
  -d '{
    "model": "higgsfield-genjutsu",
    "prompt": "The mascot dances exactly like the dancer in the source video.",
    "video_url": "https://example.com/dance.mp4",
    "reference_image_urls": ["https://example.com/mascot-front.png", "https://example.com/mascot-side.png"],
    "resolution": "720p",
    "duration": 8,
    "mode": "async"
  }\

Rights and QA

Use a dance clip you made or have the rights to. The model transfers motion, and the source dancer's choreography may belong to someone else. After generation, extract a few stills with Video frames to check that the mascot did not change shape between poses.