Guides
Seedance video
Make video from a prompt, a starting frame or up to 30 reference images, in the Station or through the API.
seedance-2.5 is the premium video model. Give it a prompt and it generates a clip; give it a starting frame or a few reference images and it animates from them, holding your subject consistent across cuts. Clips can run up to 30 seconds with synchronized audio.
seedance-2.0 works the same way at a lower price: every mode and size below applies, with clips up to 15 seconds and up to 9 reference images.
Start from a strong input
The biggest lever on quality is what you feed in. A detailed first frame or reference image gives the model real structure to animate, so the result looks the way you intended. A plain text prompt works well for open, generic scenes, but when you want a specific character or look, bring an image.
Three ways to make a clip
The mode is chosen automatically by how many reference images you attach:
- No references (text to video). Describe the whole scene. Best for open, cinematic shots where any subject will do.
- One reference (image to video). Your image becomes the first frame. Describe the motion, not the scene, since the scene is already set.
- Two or more references, up to 30 (reference to video). The images are subject and style references rather than a starting frame. Cite them in the prompt to say what each one is for.
Citing references with @image
When you attach more than one reference, refer to each in the prompt by its number: @image1, @image2, and so on, in the order you added them.
"@image1 walks through the alley lit like @image2, slow dolly in"
In the Station, each reference shows as a tile named @image1, @image2 and so on; type that name in your prompt. In the API, send the images as input_reference and cite them the same way. Order matters: the first image you attach is @image1.
Reference videos and audio are cited the same way, @video1 and @audio1, and need it most: say what to take from each, or the model may follow the images alone and leave the video unused.
"@image1 does the dance from @video1, same camera move, studio lighting"
In the Station
Open the Station, pick Video, and choose Seedance 2.5 under Model:
- Add references with +, up to 30, from your device or your library. Each becomes a tile,
@image1and so on. - Type a tile's name in your prompt to cite it.
- Set the length and size on the composer's chips. The price updates as you change them.
Sound is generated with the clip. For a silent one, turn Sound off in the composer, or turn it down later with Edit clip.
In the API
Seedance runs on the async /v1/videos endpoint like every video model: submit a job, poll until it finishes, then download the MP4. Pass model: seedance-2.5, a size and seconds, and send references as multipart input_reference fields. See Videos for the full request and polling flow.
First and last frame
With a single starting image you can also set the last frame, and Seedance animates the transition between them. It is the cleanest way to get a controlled morph, or to continue a clip: feed the previous clip's final frame as the first and a target frame as the last.
In the Station, once one picture is added an End frame chip appears; add the last frame there. In the API, send a second multipart field, last_reference, alongside the single input_reference (the first frame):
A last frame works only with exactly one input_reference (the first frame); sending it with none or with several references returns 400.
Start from a video
Seedance also works from a VIDEO, three ways:
- Reference: a new clip that borrows the video's subject, style or motion.
- Edit: the same clip changed the way your prompt says (a new outfit, a different scene, a new mood). It keeps the video's own length.
- Extend: the clip continued from where it ends, for the duration you pick.
In the Station, add an MP4 or MOV with + and choose Reference, Edit or Extend on the chip that appears with it. To keep going from something you just made, press Use as reference on the finished clip: nothing is re-uploaded. Up to 10 videos and 30 seconds in total on seedance-2.5, 3 and 15 seconds on seedance-2.0. Videos you upload are deleted as soon as your clip is made. Only use videos you have the rights to, and footage of real people only with their consent.
In the API, pass reference_videos and task (Videos).
Reference audio
Seedance can follow a voice or a piece of music. In the Station, add MP3 or WAV clips with +; in the API, pass them as reference_audios, as links or uploads. Cite each in the prompt as @audio1, @audio2, saying what it is for. Up to 10 clips and 30 seconds in total on seedance-2.5, where audio alone is enough; 3 clips and 15 seconds on seedance-2.0, which needs an image or a video too. It costs nothing extra. Use only audio you have the rights to, and a real person's voice only with their consent.
Draft first, finish in 1080p
seedance-2.5 can make a quick draft: the shot renders at 480p, at the 480p price. If it is the one, finish the same shot in 1080p: the same prompt, inputs, length, shape and sound, rendered again at full quality. Try a few drafts, then pay for 1080p only on the shot that works.
In the Station, turn on Draft and run; on the finished draft press Finish in 1080p, which shows the price before you confirm. A draft can be finished for about 2 days. In the API, send "draft": true, then {"finalize": "<draft id>"} (Drafts). Drafts work from text, images and reference videos, but not with edit or extend.
To finish a draft later, our video provider holds its prompt and inputs, and is instructed to delete them after 48 hours; Roteo stores none of it. The Station asks you to approve that on every draft run. Leave draft off for anything you do not want kept. See Privacy.
Duration, size, and cost
- Duration (
seconds): 5, 8, 12, 16, 20, 24, or 30 onseedance-2.5; 5, 8, 10, 12, or 15 onseedance-2.0. - Size: 480p, 720p or 1080p in
16:9,9:16,4:3,3:4,1:1or21:9, e.g.1280x720(720p 16:9) or960x960(720p 1:1). The full list is under Sizes. - Cost is per delivered clip and scales with duration and size, so a long clip at the largest size costs the most. A reference video adds its own length to the price (below), and the exact figure shows before you run. A failed job is free. Live pricing is in the dashboard.
What a reference video adds
The model reads every frame of a reference video, so its length is billed as input, at the clip's size, on top of the clip you get. While a video is attached, every second, input and output alike, costs about 40% less than in a clip without one. So the total depends on how long the reference is:
Prices at the time of writing; the same proportions hold at 720p and 1080p.
- A short reference is almost free, a long one is not. A 3-second reference costs less than none at all, while a 10-second one nearly doubles the price. Trim a reference to the part you need: the moves of a dance, one camera move.
- An edit runs as long as its input. Editing a 10-second clip bills a 10-second clip plus the 10 seconds it reads ($1.231 on
seedance-2.0at 480p, against $1.004 for a plain 10-second clip). - Several videos add up. Their lengths are summed: up to 30 seconds on
seedance-2.5, 15 onseedance-2.0.seedance-2.5counts each video's length in whole seconds, rounded up. - The exact price shows before you run. Station prices the run the moment a video is attached; in the API,
POST /v1/videos/quotemeasures the video and returns the price for free.
Writing good prompts
- Detail is free. The bill depends on the clip's duration and size (and any reference video's length), never on how long the prompt is.
- With a first frame, describe the motion, not the scene. The frame already fixes the look.
- For a multi-scene story, script the sequence. One clip can carry cuts, camera moves (push in, pull back, defocus), transitions, ambient sound, even an end-card text reveal. Write them in order and cite
@image1,@image2where each reference belongs. - Use rich reference images. A detailed photo animates well; a flat logo or solid color gives the model nothing to work with and can stall.