Reference-to-video is the mode. Multi-reference is the ceiling of that mode: how many stills the model will actually hold. Seedance 2.0 is the current example in the catalog, with a published stack up to 12 references and a 15-second cap. Extra images past a model's limit are dropped, not blended. Three angles of one product teach more than three near-identical crops. A face, a garment, and a location in the same stack is how a shot stays on-brand without locking the first frame.
The distinction from first-last frame and from image-to-video matters. An image-to-video input is the opening picture; its framing is the shot. First-last frame is two stills that bound a motion. A reference is not a frame at all. It is an identity the model may place at another angle, under other light, at another size. Several of those at once is how a recurring subject survives a new camera move.
Everything in a reference leaks. A product on a red sofa brings the sofa. Clean plates and cutouts carry less of what you did not intend. The prompt still owns the action. The stack owns who and what is in the shot. Any-to-any models fold this into a broader input list (text, image, video, audio); multi-reference is the stills-only version of that idea, and it is the one most creator stacks actually run today.
In practice
- Check the model's reference limit before you add the fifth still. The extras will not be used.
- Disagree on purpose: different angles beat duplicates.
- Crop out the background you do not want inherited. The model cannot tell which part you meant.
The mistake to avoid
Submitting a moodboard of twelve JPEGs to a two-reference model and blaming the prompt when ten of them vanished.
Where you will run into it
- AI Video Generator — Text-to-video, image-to-video, and story chaining in one studio.
- Seedance 2.0 — Seedance 2.0 generates Hollywood-grade cinematic videos from text prompts with native audio-visual synchronization, director-level camera and lighting control, and exceptional motion stability
Related terms
Reference-to-video
Reference-to-video means the model builds a clip from separate images of a subject it can place anywhere. Those images are an identity, not the first frame.
Reference image
Reference image meaning: a picture supplied with the prompt so the model can copy identity, product, or style without using it as a frame.
Character consistency
Character consistency meaning: whether the same person, mascot, or product still looks like itself across separate generations.
Any-to-any
Any-to-any is a model that takes mixed inputs — text, image, audio, video — and returns mixed outputs, rather than a one-way image-to-video pipe.
Text-to-video
Text-to-video is generation from a written prompt alone — you describe a shot, the model invents every frame of it, and no image or footage goes in.
The all-in-one AI studio for creators. 60+ models for video, image, voice, music and lipsync, in your browser or on your phone.
Free on the web and on iPhone, one account.