Reference video or motion transfer: Seedance 2.5 or Genjutsu?

Send a reference video to Seedance 2.5 when you want a new clip inspired by it, and use Genjutsu Motion Transfer when you want the source clip's own length and framing kept while the subject changes. ByteDance describes Seedance 2.5 as understanding a reference video's intention, framing and cinematic language, going beyond motion transfer into creative interpretation. On Sume, seedance-2.5 takes reference_video_urls and writes a fresh clip; higgsfield-genjutsu takes one source video_url plus 1 to 8 images and keeps the source's length and framing.

What is the difference in one table?

Both accept a video, but they use it differently. The Video Router keeps two fields apart: video_url is a source clip, used for the Gemini Omni edit and the person-swap models, and reference_video_urls conditions a fresh generation.

What does ByteDance mean by going beyond motion transfer?

On the Seedance 2.5 product page, the Seed team says the model "understands reference videos more precisely", capturing the intention, framing and cinematic language, not only the motion. The launch post shows reference prompts such as continuing a clip with the same characters and sound, and editing a clip while keeping characters and actions and changing only the camera path.

That is a statement about the model. On Sume the Seedance rows do not expose a video_url edit: the Video Router doc lists the video_url edit mode under Gemini Omni Flash 1.1, and the seedance-2.5 row reports no video-to-video capability. So the Seedance route is reference-to-video: you describe the new clip and attach the video as guidance.

Which one should you pick?

Pick Seedance 2.5 when the reference is a mood, a camera move or a pacing example and the output length is your choice. Pick Genjutsu when the source is the performance you want to keep and only the subject changes. Pick H3 Max Recast when the job is swapping people in the source: it takes one photo per person, 1 to 4 photos, at 768p or 1080p.

Check a source clip's length before a Genjutsu run. Its duration must equal the inspected input length rounded up, within 4 to 30 seconds, so a 3-second source cannot be used.

What does the vendor say the references can hold?

The launch post says Seedance 2.5 accepts up to 30 images, 10 video clips and 10 audio clips as references in one pass, which is the vendor's figure for its own platforms. Sume's per-model limits differ, so read the capabilities on GET /v1/video-router/models for the numbers your request must fit instead of copying the vendor's totals.

The reference-video use cases in the post are concrete: extend a clip with the same characters and sound, or take a clip and change only the camera path. Both are written as prompts that say what to keep and what to change, and that wording carries over to a Sume request.

How do you send each request?

Through the Video Router both are one POST /v1/video-router/generate with a different model. For Seedance, add reference_video_urls and a prompt that says what to take from it. For Genjutsu, send video_url, reference_image_urls and a duration matching the source. Do not combine first/last frame fields with reference fields on Seedance; the request is refused. If the reference audio matters, add it, but reference_audio_urls needs at least one reference image or video alongside.

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: ref-video-001" \
  -d '{
    "model": "seedance-2.5",
    "prompt": "Match the slow orbit and pacing of the reference; new subject: a red kettle on a stove",
    "reference_video_urls": ["https://example.com/orbit-reference.mp4"],
    "duration": 10,
    "resolution": "480p",
    "mode": "async"
  }'