Higgsfield Genjutsu API on Sume: one source video, 1 to 8 images

Genjutsu Motion Transfer on Sume is not a text-to-video model. Every request needs one source video and between 1 and 8 reference images; a body with no video, two videos, or no images does not fit the model's input shape.

The shapes below are from Sume's catalog notes and Video Router docs, read 2026-09-30. The model is listed only when its provider is configured, so check the catalog first.

What does the request look like on each route?

The same inputs go in different fields depending on the route you call.

{
  "model": "<catalog id from GET /v1/videos/models>",
  "input_references": [
    { "type": "video_url", "video_url": { "url": "https://media.sume.com/artifacts/artf_demo/move.mp4" } },
    { "type": "image_url", "image_url": { "url": "https://media.sume.com/artifacts/artf_demo/character.png" } }
  ]
}

Why is a text-only prompt not enough?

The catalog marks text-to-video and image-to-video as false for this model. It transfers the motion in the source video onto the characters in your images, so both inputs are required.

What else is fixed by the source video?

Output keeps the source length and framing. There is no aspect_ratio, frame_images, generate_audio, bitrate_mode, or audio references. The catalog says duration must be the inspected input video's length in seconds, rounded up, within 4 to 30. See how to inspect the source duration first.

How many images can I send?

Up to 8 on Sume. Higgsfield's own limits are covered in Genjutsu reference images: 30 on Higgsfield, 8 on Sume.

How do I get the exact id and fields?

Call GET /v1/videos/models and read the entry for the Motion Transfer model, including supported_input_references. If it is not listed, its provider is not configured for your deployment. The video docs describe the catalog fields.