How many references a model will take, and what they cost
Keeping a face the same across separate generations means supplying references, and every model caps how many it will take in one call. The caps range from one to thirty, and one vendor charges beyond the fifth. As of 2026-09-12.
| Model | Image references per call | What the vendor adds |
|---|---|---|
| MiniMax H3 | Up to nine | First five free, then $0.04 each |
| Seedance 2.0 and 2.5 | One to nine, one to thirty | References with real human faces are not accepted |
| Vidu reference-to-video | One to seven | One to four on q2-pro when videos are uploaded |
| Wan 3.0 | Up to ten | Addressed as Image 1, Image 2, in the order supplied |
Inclusion rule. Models whose vendor publishes a reference count on an API page. A model asserting consistency without publishing a cap is not given a row. Order. Alphabetical by model name.
1A cap is a design constraint, not a detail
A production with six recurring characters and a cap of four cannot put the whole cast in one call. That forces either fewer characters per shot or a different model, and it is better discovered at the shot-list stage than at the first crowd scene.
The caps also differ between tiers of the same model, so a cheaper tier can quietly hold fewer references than the one the tests were run on.
2Ordering is part of the contract
One vendor documents that references are addressed by position. That makes the order in the request part of the prompt, and it means a library that reorders its exports will silently change what the model was asked for.
3Per-reference pricing changes the arithmetic
Where references past the fifth are billed, a six-character shot costs measurably more than a five-character one on the same second of video. That is the only place on this register where the cast size appears directly on the invoice.
4References do not fix wardrobe or lighting
A face reference keeps a face. It does not keep the jacket, the haircut of a later episode, or the way a room lights that face. Productions that treat references as a complete continuity system find the face steady and everything around it moving.
The rest has to come from the shot text and from separate references, which is why the cap on how many references a call takes is a production constraint rather than a technical footnote.
5Consistency degrades quietly, not suddenly
Nothing announces that a character has drifted. Each generation is plausible on its own and the change is only visible across a cut, which is why drift is usually found in the edit rather than at generation time.
Checking each new shot against the reference rather than against the previous shot is the habit that catches it, because comparing against the previous shot compares against something that may itself have drifted.
6Where these figures come from
A definition kept for this register: the term, and what vendors state about it, each figure read on 2026-09-12. The tools themselves are on the register. Related: Storyboard script, Scene plate.
7How a serial stays consistent
Serial work like “Character consistency” depends on continuity; SceneMixer's guides describe how it keeps it:
That is SceneMixer's core design: characters live in a persistent asset library with locked reference art, and every segment that features a character is generated against those references — same face in episode 1 and episode 50.
SceneMixer, The AI Micro Drama Production Pipeline — “Can AI keep the same actors across 50 episodes?” (read 2026-10-13)
The cast is generated once, at the start, as full-body reference sheets. Every segment that features a character is generated against that character's sheet, with face and body crops attached as references — so the face in episode 1 is the face in episode 40.
SceneMixer, AI Short Drama Generator — “How does it keep the same actor across episodes?” (read 2026-10-13)