Hand the model a cast list instead of a first frame. Upload photos of people, products, outfits or places, tag them in your prompt, and Yovid stages a new scene around them on Seedance 2.5, Wan 3.0, MiniMax H3 and more.
A reference is something the model has to recognise, not a picture the clip has to begin on. Yovid passes each file to the model as a source of identity, wardrobe, product detail, setting or style, and the model composes a fresh shot with the camera and action you write. That is how one character carries a whole series, or a product photographed on a white table turns up in a street scene.
Use the picture as the first frame and keep its composition, framing and pose, with motion added.
Image to video AIPin both the opening and the closing keyframe and let the model fill in the move between them.
First and last frame to videoUpload the clip and swap the product, the background or the look while the camera move stays.
Video to video AISeedance 2.5 reads up to 30 reference images and 10 reference clips per run in the Yovid composer. ByteDance’s demo below uses eight: a paper puppy’s character sheet, a real Labrador, a bedroom, a pet feeder, a toy box and a child’s drawing, all cast in one 25-second scene.
Seedance 2.5Wan 3.0 takes up to 10 images and 5 clips. In Alibaba’s example one portrait fixes the heroine, one image fixes the lotus realm, and seven sheets each show a creature in 2D and 3D; all nine appear in the same 30-second take.
Wan 3.0The two faces come from portraits, the slurping action from a short clip. The Seedance 2.0 models take up to 9 images and 3 clips of 2 to 15 seconds per run; tag each file as @Image or @Video so the model knows what it is for.
Seedance 2.0A mood board, a character sheet, a handbag shot from three angles and the closing logo: four references let MiniMax H3 cut a 15-second 2K campaign in which the bag and the end card stay on-model.
MiniMax H3Three photos of one pizza maker and his shop kept him recognisable through a night view from the street, the oven and the hand-off at the counter in this Yovid run. Reuse the same set and the same wording to build a series around him.
Seedance 2.0Seedance 2.5 holds the largest stack, Wan 3.0 runs 30-second takes from up to 10 images, MiniMax H3 renders at 2K, and Seedance 2.0 Mini lets you test the workflow on free credits.
Limits and rates come straight from the generator’s own settings, so this table changes when a model does.
| Model | Reference images | Reference clips | Clip length | Output length | Top resolution | On free credits | Credits per second |
|---|---|---|---|---|---|---|---|
 Seedance 2.5 | 30 | 10 | 2–30 s | 4–30 s | 1080p | No | 6.1 at 720p |
| Seedance 2.0 | 9 | 3 | 2–15 s | 4–12 s | 1080p | No | 4 at 720p |
| Seedance 2.0 Fast | 9 | 3 | 2–15 s | 4–15 s | 1080p | No | 3.75 at 720p |
| Seedance 2.0 Mini | 9 | 3 | 2–15 s | 4–15 s | 720p | Yes | 2.5 at 720p |
| Wan 3.0 | 10 | 5 | 1–15 s | 2–30 s | 1080p | No | 2.4 at 720p |
 MiniMax H3 | 9 | 3 | 2–15 s | 5–15 s | 2K | No | 4.55 at 2K |
 Kling O3 | No | 1 | 3–10 s | 4–10 s | 1080p | No | No |
Credits per second are catalogue rates for a run with image references, before plan discounts. Images add nothing on the Seedance models or Wan 3.0; MiniMax H3 adds about 1.4 credits for each image beyond five. A reference clip changes the bill: Seedance bills clip plus output seconds at its cheaper reference-clip rate, Wan 3.0 and MiniMax H3 bill both at the normal rate, and Kling O3 bills the output alone. On Wan 3.0 the clips and the output share a 30-second budget. Free credits stop at 720p.
Use cases
Keep the same host, mascot or character through every episode, ad or short by reusing one reference set.
Put a packshot into a kitchen, a trail or a city street while the shape, colour and logo stay true.
Combine models, garments and accessories from separate photos into one styled fashion film.
Character sheets and drawings become animated performers that keep their design from shot to shot.
Step 1
One image per character, one for the outfit or product, one for the place or style. A plain background and a clear view help most.
Step 2
Add three or more images and they become references. Typing @ in the prompt lists your uploads as @Image 1, @Image 2 and onward; give each its role.
Step 3
Write the action, camera and light of a scene that does not exist yet, pick a length, check the credits and press Generate.
“The face from @Image 1, the jacket from @Image 2” beats five photos and no instructions. Name every file’s role in the prompt.
A front, side and back view, even combined in one sheet, keeps a person or mascot recognisable while the camera moves around them.
Skip the written description of the face or outfit. Words that disagree with the image pull the result away from it.
Keep the images, the tags and the phrasing identical from prompt to prompt, and let only the action and the camera vary.
A plain background and an instruction such as “keep the label and cap exactly as in @Image 3” protect the details that sell.
Model providers may refuse celebrities, real people’s faces or protected characters. Failed generations are not charged.
Related tools
References are one input among several. Pick whichever matches the material you have, with the same account and credits.
Prompt guideSeedance 2.5 prompt guideHow to structure prompts that tag images and clips by role.Read guide
Model comparisonSeedance 2.5 vs Seedance 2.0Reference limits, length and edit modes, side by side.Read guide
Model comparisonAI video models comparedFrames, references, length and resolution for 14 models.Read guide