Reference to Video
Lock a character or a style with reference assets and hold it across shots. Seedance 2.5 accepts up to 30 references in a single request.
Pay as you go, with the exact cost shown before you generate
Pass in character, scene or style references and the model keeps them consistent while it generates — the capability serialised and episodic work depends on. Seedance 2.5 takes up to 30 reference images, 10 reference videos and 10 reference audio clips per request, and the prompt can point at a specific asset by position with @image1 or @video1. Seedance 2.0 allows 9 and 3 respectively.
- 01
Upload character, scene or style references
- 02
Point at a specific one from the prompt with @image1
- 03
Check the credit cost and submit
Real generated samples
Every clip below was generated on this platform. The prompt is reproduced as written.
Majestic snow-capped peaks above a flowing sea of clouds at sunrise, cinematic aerial shot.
Retro gaming products on a dark desk, premium commercial motion and controlled studio light.
Aerial view of a modern city at golden hour, cinematic movement and photorealistic detail.
Cost per second
1 credit = $0.01. The table shows 480P. Higher resolutions and longer clips cost more, and the exact figure is shown before you generate.
| Model | Per second | Max length | Resolutions |
|---|---|---|---|
| Seedance 2.5 | $0.165 | 30s | 480P / 720P / 1080P |
| Seedance 2.0 | $0.113 | 15s | 480P / 720P / 1080P / 4K |
| Seedance 2.0 Fast | $0.075 | 15s | 480P / 720P |
| Seedance 2.0 Mini | $0.058 | 15s | 480P / 720P |
In depth
Three jobs reference mode is built for
Most work that needs reference mode falls into one of three patterns. They differ in what each attachment is for, and saying that out loud in the prompt is most of the craft.
- Story series and short dramas with a recurring cast
- The same people appear episode after episode, often alongside a product or prop. Give each reference one job and state it: Image 1 is the main character only, Image 2 is the product only. A photorealistic 3D character can be refused as a real face just like a photo; saving the character under Assets first is the same fix, and it keeps the look identical across the series.
- Motion-reference demos: sports drills, exercises, dance
- Attach a clip of the movement as @video1 and the character as @image1, then say which does what: the video supplies motion only, the image supplies identity. BytePlus documents this pairing as motion plus subject reference. One movement per clip in a single continuous take tends to hold together better than a combination. The reference clip itself must be at least 2 seconds long.
- Storyboards to shots for games and short films
- A multi-panel storyboard in one image is read as a plot reference, not a frame-by-frame spec. BytePlus recommends 15 panels or fewer, simple line art and no text on the panels, with actions and camera movement written into the prompt. Attach characters separately and label the storyboard as environment and plot only. If a shot has to start and end on exact frames, use first and last frame in image-to-video instead.
How references are addressed in the prompt
Reference-to-video is the only mode where the prompt and the attachments have to agree with each other. Attached assets are numbered in the order you select them, and the prompt refers to them by that number — the workbench shows tags like @Character1 and @Character2, which are translated into Image 1 and Image 2 when the request is built.
Two consequences worth knowing. Reordering your selection renumbers everything, and the workbench rewrites the tags in your prompt to match, so a reference never silently points at the wrong asset. And you should never paste a raw asset ID into the prompt — the model does not resolve IDs, and the official guidance is explicit about it.
Placement matters more than people expect. Put the tag next to the thing it describes — "@Character1 walks into the rain" anchors the character to that action, while listing all tags at the end of the prompt leaves the model to guess which reference belongs to which subject.
How much you can attach
Seedance 2.5 accepts up to 30 reference images, 10 reference videos and 10 audio clips in a single request. The Seedance 2.0 family accepts 9, 3 and 3 respectively. Exceeding a limit is rejected before the request reaches the model, so nothing is charged.
A reference video has to be at least 2 seconds long, and no single clip may exceed 30 seconds on Seedance 2.5 or 15 on the 2.0 family. This is the reference clip's own length — it has nothing to do with how long your output will be. A 1.5-second reference fails even when you are only asking for a five-second result.
Audio-only generation — no image and no video attached — works on Seedance 2.5 only. On the 2.0 family, audio has to accompany at least one image or video reference.
Keeping a character consistent across shots
This is what reference mode is actually for. Generating three clips from three similar prompts gives you three different-looking characters; attaching the same reference asset to all three gives you one character in three shots. The reference does the identity work so the prompt can be about action and camera.
Assets can be deleted, and a reference to a deleted asset fails upstream — retrying will never succeed, because the asset genuinely no longer exists. Reopening the character picker clears stale selections automatically, so the fix is to select again rather than to retry.
Questions
- How do I point the prompt at a specific reference?
- Attached assets are numbered in selection order and the prompt refers to them by number — the workbench shows tags like @Character1, which become Image 1 in the request. Put the tag next to the subject it describes rather than listing all tags at the end. Never paste a raw asset ID into the prompt; the model does not resolve IDs.
- How long does a reference video have to be?
- At least 2 seconds, and no more than 30 seconds per clip on Seedance 2.5. This is the reference clip's own length, not the length of the video you are generating.
- Can a video drive the motion while an image sets the character?
- Yes. Attach the movement as @video1 and the character as @image1, and state in the prompt that the video supplies motion only and the image supplies identity. It suits sports drills, exercise demonstrations and dance.
- How are credits consumed?
- All video generations consume credits based on the selected model, duration, and quality. You can see the estimated cost before generating and top up credits with Stripe when needed.
- Are credits refunded when generation fails?
- Yes. If a generation fails because of a system, provider, or task error, frozen credits are released automatically. Review the task in My Creations and contact support if anything looks incorrect.
- Can I use the videos commercially?
- Yes, you own the commercial rights to the videos generated via AiuniVid. You can use them for social media, ads, YouTube, etc. Please refer to our Terms of Service for details.
- How do the Seedance versions differ?
- Seedance 2.5 supports omni references, audio, and clips up to 30 seconds. Seedance 2.0 is best for high-quality visuals and higher resolutions. Fast is quicker, and Mini is lower-cost for drafts and batch tests.
- What if generation fails?
- If generation fails due to a system error, credits are refunded automatically. For help, message us on Telegram @Jarvis11x or email [email protected].
