Video to video motion transfer with AI: MiniMax H3 reference video

For video-to-video motion transfer with MiniMax H3 on Sume, send a motion video and one or more character images in input_references on POST /v1/videos, and tell the prompt which reference does what: for example "Use Video 1 as the motion reference for a street-dance performance. Use Images 1 and 2 as the character references." That prompt is from fal's guide; MiniMax's announcement lists V2V motion transfer among H3's strengths.

Vendor facts are from MiniMax's announcement and fal's prompting guide; limits from the Sume Video generation docs and code, read 2026-09-29.

How do I send it?

Put the motion clip and the characters in one input_references array. The prompt names them by order: Video 1, Image 1, Image 2.

curl -X POST https://api.sume.com/v1/videos \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: h3-motion-001" \
  -d '{
    "model": "minimax-h3",
    "prompt": "Use Video 1 as the motion reference for a street-dance performance. Use Images 1 and 2 as the character references.",
    "input_references": [
      { "type": "video_url", "video_url": { "url": "https://example.com/dance.mp4" } },
      { "type": "image_url", "image_url": { "url": "https://example.com/dancer-a.png" } },
      { "type": "image_url", "image_url": { "url": "https://example.com/dancer-b.png" } }
    ],
    "resolution": "768p",
    "aspect_ratio": "9:16",
    "duration": 10
  }'

What limits apply?

Is there a dedicated motion transfer endpoint on Sume?

Yes, for a different model: Kling 3.0 Motion Control at POST /v1/kling/3.0/motion-control animates one still with a driving video and follows the driving clip's length; see Motion control API. The H3 route above is reference-to-video, where the prompt, not a dedicated field, says which clip carries the motion.

How closely will the motion match?

Nothing in the pages read gives a fidelity figure. fal's guide says to assign each reference a specific job; that is the lever you have. Generate a short clip first, compare it to the source, and adjust the prompt wording before a longer one.