Keep one look across a video series with reference images

To keep one visual look across a series of AI videos, send the same reference image in input_references with every request to Sume's POST /v1/videos, and change only the prompt. That puts the look in the request instead of in each prompt's wording. Higgsfield's Genjutsu Restyle, per its 2026-09-30 changelog entry, redraws a whole video in one of 20+ styles; a reference image is how you pin a style when you generate new clips instead of restyling existing ones.

Request fields come from Sume's Video generation docs, read 2026-10-02.

How does input_references work?

input_references provides style or content reference images for reference-to-video: the model uses them as visual guidance, not exact frames. frame_images is the other route, for first and last frames, and if both are present frame_images wins and the request becomes image-to-video. Only a model whose supported_input_references lists image_url accepts it, so read that field from GET /v1/videos/models.

{
  "model": "seedance-2",
  "prompt": "Product on a desk, slow push-in, same look as the reference",
  "input_references": [
    {"type": "image_url",
     "image_url": {"url": "https://example.com/look-reference.png"}}
  ],
  "resolution": "720p",
  "aspect_ratio": "9:16"
}

What should stay the same across the series?

Keep the reference image, model, resolution, and aspect_ratio fixed. Change one thing per clip: the action in the prompt. Put the clip-specific direction first and the shared style sentence last, the same in every request.

When is this the wrong tool?

If you already have footage whose motion you want to keep, reference images do not carry that motion. For that, use a video-to-video route; three ways to restyle a clip compares them.

How do I get the results?

The create call returns a job id and polling_url; poll until completed, then download from unsigned_urls[0], or pass callback_url (HTTPS) to be notified.

Pick the reference image deliberately: a frame with the palette, lighting, and texture you want, and without a subject you do not want repeated. Test it on two clips before running the series. If the look drifts anyway, add the same short style sentence to every prompt and compare, changing one of the two at a time.