Reference to video
The mode that solves the hardest problem in AI video: the same face, the same product, the same look, shot after shot. Upload up to nine references, call them out in the prompt as Image 1 and Image 2, and direct the scene around them.
Made with this model
Every clip below was rendered here, on the settings shown under it. The caption is the exact prompt that produced it.
Seedance 2.0 · 720p · 16:9 · 4s
MiniMax H3 · 2K · 16:9 · 5s
How reference to video works here
Upload your references
Up to nine images: a character from a couple of angles, a product on its own, a colour treatment you want carried across — anything that has to stay recognisable.
Reference them by number
“Image 1 is the barista, Image 2 is the cup.” The numbering is how the model knows which reference belongs to which part of the scene.
Direct the scene
Everything else is an ordinary prompt: action, camera, light. The references constrain identity; the prompt decides what happens.
Specs at a glance
| Model | Duration | Resolution | Native audio | Reference images | From |
|---|---|---|---|---|---|
| Seedance 2.0 | 4–15 seconds | 480p · 720p | Yes | Up to 9 | 31 credits for 5 seconds |
| MiniMax H3 | 5–15 seconds | 2K | No | Up to 9 | 130 credits for 5 seconds |
Getting identity to stick
One subject per reference image. A photo containing two people gives the model an ambiguous target.
Match the lighting you are about to ask for. A hard-flash reference fights a golden-hour prompt.
Two or three angles of the same character beat nine near-identical crops.
Name every reference in the prompt. An uploaded image the prompt never mentions is largely ignored.