Seedance 2.5 Reference to Video: a walkthrough from cast to final cut

Most video models take one picture and a sentence. Seedance 2.5 Reference to Video takes a whole production folder: up to thirty pictures, up to ten video clips and up to ten audio files, plus the prompt that says what each of them is for. This guide walks through one shoot from start to finish, in the order you would do it on a real set.

The example is a thirty-second-long idea cut into short shots: a courier on a bicycle crossing a rainy city at dusk. The same steps work for a product, a character or a place.

Step 1: decide what must not change

Before you gather anything, write one line each for the things that have to stay the same in every shot: the face and clothes of the courier, the bicycle, the look of the street. Everything you attach later exists to protect one of those lines. If you cannot say what a file protects, leave it out. A crowded folder is the quickest way to get a muddy result.

Step 2: build the cast of references

• Pictures are the cast. One clear image of the courier from the front, one from the side, one of the bicycle on its own. Plain backgrounds, even light.
• A video is for movement or camera work you want to borrow: a slow tracking shot, a handheld run. The total length of all videos together must stay within 30 seconds.
• An audio file is for voice, music or the feel of the sound. The total length of all audio files must also stay within 30 seconds.

Pictures are the only input that is on by default in the form. Videos and audio sit under the advanced options, and you need at least one reference of any kind for the model to work as a reference model. With none, the request runs as plain text to video.

Step 3: write the prompt as a call sheet

The prompt names each file in the order you uploaded it, as @Image 1, @Video 1, @Audio 1, and says what it is for. Then it describes the scene, the action, the camera and the mood. Think of a call sheet: who is in the shot, what they do, where the camera stands.

Step 4: pick the settings

• Duration runs from 4 to 30 seconds. Shorter clips hold a likeness better. For a longer film, cut it into shots.
• Resolution offers 480p, 720p, 1080p and 4K. Draft at 480p or 720p, then run the keeper again at 1080p or 4K.
• Aspect ratio offers 16:9, 9:16, 4:3, 3:4, 1:1 and 21:9. Choose it before you run, since framing is part of how the shot is built.
• Generate audio is on by default and gives the clip native sound that follows the picture.

The cost shown on the model page is the figure for the cheapest setup, and it rises with length and resolution. Check it before a long or large run, and spend drafts at the small sizes.

Step 5: review like an editor

1. 01Watch once with the sound off. Is it the same person and the same bicycle from the first frame to the last?
2. 02Watch once with the sound on. Does the sound match what the picture shows?
3. 03Name the one thing that is wrong, and change only that. A different file, or one sentence.
4. 04When it is right, rerun the same inputs at the final size.

For the thinking behind good references, reference images for a consistent style is the longer read. If you want to move a character through someone else's performance rather than keep a look, see the Motion Control guide. To stretch a story past one clip, make a longer video from matching clips.

What to expect, honestly

A reference is a strong hint, not a guarantee. Faces and logos can still drift in a long clip, hands in busy action are the usual weak spot, and a prompt that fights its own references will lose to one of them. Keep clips short, keep the reference set small, and treat the first result as a draft.