Keep the character consistent across the clip
Upload a first frame and up to 30 character references, and condition the generation on them. Consistency is best-effort rather than identity lock, output is 480p or 720p, the credit estimate appears before you generate, and you need an account to generate.
このページは英語で提供しています。ワークスペース・料金・FAQ は日本語で利用できます。
- Character consistency here means conditioning the generation on reference images, not locking an identity.
- Seedance 2.5 accepts up to 30 image references, plus 10 video and 10 audio references.
- The first-and-last-frame mode takes one or two images and controls where the shot starts and ends.
- Every generation is independent: the same character across several shots means reusing the same references.
Reference limits at a glance
What does character consistency actually mean?
It means the model is conditioned on reference images of the character, so the face, hairstyle and clothing it renders lean toward that reference. It is not an identity lock, a face swap or a guarantee.
Treat every generation as a fresh draw with strong guidance. Two runs from the same references will look like the same person most of the time, but small details — earrings, a hair parting, a logo on a jacket — can shift. That is why the workflow below ends with inspection rather than trust.
How many references can you supply?
It depends on the model, and the ceiling is enforced at generation time rather than in the UI only.
| Model | Image references | Video / audio references | Note |
|---|---|---|---|
| Seedance 2.5 | 30 | 10 video, 10 audio | Default selection; longest single take at 30 seconds |
| Seedance 2.0 (omni reference) | 9 | — | Use when the delivery has to be 4K |
| Other models in the workspace | 1 first frame | — | Image-to-video conditioning without the omni reference set |
How do you prepare reference images?
Reference quality decides the result more than the prompt does.
- 01
Shoot or select three to five clear angles
Front, three-quarter and profile cover the angles a clip usually needs. Extra images beyond that help only if they add a genuinely new view.
- 02
Keep lighting consistent across references
Mixing a hard sunlit photo with a soft studio one teaches the model two different faces. Prefer one lighting setup for the whole set.
- 03
Avoid occlusions and cropping
Hair across the face, a hand over the jaw or a tight crop removes the information the model needs. Full head and shoulders work best.
- 04
Write the motion, not the identity
Describe what the character does — turns, lifts, speaks, walks — and let the references carry who they are.
What still breaks consistency?
The failure modes are predictable, and most of them are visible before you spend credits.
- Extreme angles: a profile or a top-down view is rarely in the reference set, so the model invents it.
- Fast motion and large occlusion: a hand crossing the face or a quick turn loses the features.
- Wardrobe changes between shots: keeping the same outfit across shots needs the same references, not a rewritten prompt.
- Multiple characters in one generation: the model blends features when two identities compete for the same reference budget.
- Low-resolution references: below roughly 700 px on the long edge, detail is already gone before generation starts.
How do you keep one character across several shots?
There is no cross-generation memory, so consistency across shots comes from reusing the same reference set and stitching the results.
- Save the reference set as a folder and attach the same images to every shot.
- Use the first-and-last-frame mode when a shot has to end on a composition that the next shot starts from.
- Generate each shot separately, then assemble them in the long video mode, which supports up to 180 seconds.
- Label the shots by scene and keep the prompt template identical apart from the action.
Character consistency questions
Can this keep the same face across a clip?
It conditions the generation on your reference images, which keeps the face, hair and clothing close to the reference in most cases. It is not an identity guarantee, so inspect the result before publishing.
How many reference images can I upload?
Seedance 2.5 accepts up to 30 image references, plus up to 10 video and 10 audio references. Seedance 2.0's omni reference accepts up to 9 images. Other image-to-video models take a single first frame.
Do I need a first frame as well?
Not required, but it anchors composition. The first-and-last-frame mode accepts one or two images and defines where the shot begins and ends.
Which model should I choose?
Seedance 2.5 is the default and gives the longest single take at 30 seconds with native audio. Switch to Seedance 2.0 only when the delivery has to be 4K.
What is the maximum clip length?
4 to 30 seconds per generation on Seedance 2.5, selectable per second. Sequences longer than that are assembled in the long video mode, up to 180 seconds.
What resolution and cost should I expect?
480p or 720p on Seedance 2.5, charged at 15 credits per second at 480p and 30 per second at 720p. A 10 second 480p clip is 150 credits.
Can I use photos of a real person?
Only with that person's permission and within the model provider's terms. Uploading someone else's likeness without consent is not permitted, regardless of what the tool can technically do.
What if two characters appear in the same shot?
Results degrade because the reference budget is shared. Generate one character per clip and combine the shots afterwards, or accept that features may blend.
Can I keep the same character across three separate shots?
Yes, by reusing the exact same reference set for each generation and keeping the prompt template identical apart from the action. There is no memory between generations.
Is the audio generated for the character too?
Seedance 2.5 generates a synchronised audio track with the clip. It is generated sound rather than a specific voice you control, and it is not a licensed music library.
Key takeaways
- 1.Consistency is conditioning, not identity lock — always inspect the result.
- 2.Three to five clean, consistently lit angles beat thirty redundant photos.
- 3.Seedance 2.5 supports 30 image references; Seedance 2.0's omni reference supports 9.
- 4.Cross-shot consistency comes from reusing the same references, not from a longer prompt.