Seedance 2.5 video-to-video: a reference clip is not a person swap

Searching for "seedance 2.5 video to video" usually means one of two jobs, and Sume's catalog splits them. If you want a new clip that is guided by an existing one, seedance-2.5 accepts reference videos and generates fresh pixels for 4 to 30 seconds at 480p, 720p or 1080p. If you want the same footage with different people in it, that is an edit of a source clip, and the catalog id for it is h3-max-recast (Video Router docs, read 2026-10-03).

The distinction is whether the model is allowed to invent the shot. A reference is guidance. A source is the thing being modified.

What Seedance 2.5 does with a video input

On /v1/videos, Seedance 2.x models list video_url among supported_input_references, and the docs state that audio and video references are honored by the Seedance 2.x models, Wan 3.0, MiniMax H3 and MiniMax H3 Max. The generated clip is a new render: it can borrow camera movement or pacing from the reference, but it is not frame-locked to it, and the output length is the duration you request, not the length of your clip (Video generation docs).

That is useful when you want a new scene with the same camera move. It is the wrong tool when you need the exact performance, the exact cuts and the original voice track to survive, because none of those are guaranteed to carry over from a reference.

What Recast does instead

h3-max-recast takes video_url as the source and 1 to 4 reference_image_urls, one per person. Duration is the source length (5 to 30 seconds, no shot over 15 seconds), the prompt is optional, and the output keeps the source's motion, camera, cuts and sound. fal documents the same behavior for its Recast product and bills 0.30 dollars per second at 768p and 0.45 at 1080p (read 2026-10-03).

Choosing in one call

Do not guess from memory: read capabilities for each model from the catalog, because limits are per model. The call below lists the catalog so a script can decide. If the result shows the id you want with a video_url capability, use it; if it does not, the docs say ids not listed are rejected on generate.

curl https://api.sume.com/v1/video-router/models -H "Authorization: Bearer $SUME_API_KEY"

Cost and time before you commit

Seedance 2.5 and Recast bill very differently, which matters once you compare them for a bulk job. Sume bills every catalog model at the provider list price times 1.25, reserved when you submit and captured when the job completes, with a refund if it fails (Video Router docs). Recast's list price is per second of source video, so a 20-second source at 768p is $6 at fal's list (20 times $0.30) and $7.50 through Sume's catalog; at 1080p it is $9 at list and about $11.25 through Sume. Seedance 2.5 is priced from provider video tokens, which the docs expose through the model catalog rather than as one flat per-second number, so read the quote for your resolution before you estimate.

Time works the same way. Neither model returns instantly, so submit asynchronously and use a webhook or poll the job rather than holding a connection open. The Jobs and results guide describes the four modes, and sync waits are capped at 30 seconds.

A decision rule

Ask: must the people change while everything else stays? Use Recast. Must the camera move and mood come from my clip but the content be new? Use Seedance 2.5 with a reference video. Must a still photo perform the exact movement from my clip? That is Kling motion control, covered in Seedance reference video or Genjutsu motion transfer for the neighboring Genjutsu case. When two ids both look plausible, run a 5-second test at the lowest resolution first; a 5-second 768p Recast costs about $1.88 on Sume.