Quick answer: minimax_h3_style_preset takes six values across two axes, staging (drama vs action) and treatment (live action vs animation, with an optional photoreal pass). Presets exist only on the multimode variant, which accepts prompt-only text-to-video as well as image-to-video. An invalid value is rejected rather than ignored.
If you call H3 through the API on the 1440P lane, one parameter is easy to skip and hard to recover from: minimax_h3_style_preset. It takes six values, and the values are named after Chinese production categories that split "what the scene is doing" from "what surfaces it is made of". Skip it and you get the default; pick wrong and you get a shot that is technically fine and dramatically wrong.
Here is the list. The third column translates the workflow each value maps to, so you can reason about the preset instead of guessing from the identifier.
The six presets
| Value | Label | Upstream workflow, translated |
|---|---|---|
live_action_drama | Live action · Drama | Drama staging · live action · dialogue register |
live_action_drama_realistic | Live action · Realistic drama | Same, plus a photoreal texture pass |
live_action_action | Live action · Action | Action staging · live action · fight register |
live_action_action_realistic | Live action · Realistic action | Same, plus a photoreal texture pass |
anime_drama | Animation · Drama | Drama staging · animation · dialogue register |
anime_action | Animation · Action | Action staging · animation · fight register |
Two axes: staging (drama vs action) and treatment (live action vs animation, plus an optional photoreal-grade pass).
What the two staging registers mean
Drama staging is the dialogue-and-emotion register: characters standing, sitting, listening, turning; the camera holds or drifts; the drama is in faces and blocking. If your prompt is "she reads the letter, then looks up", this is the correct mode.
Action staging is the choreography register: contact, chase, fall, impact; the camera moves with the action and the shot is built around motion beats. If your prompt is "he vaults the railing and lands running", this is the correct mode.
Feed a dialogue beat to an action preset and you tend to get unmotivated movement — the model stages something, because staging is what the preset is tuned to produce. Feed a fight to a drama preset and you get two people standing near each other very intensely. Neither is a bug; the preset is doing its job on the wrong script.
The photoreal pass is a grading decision, not a quality tier
The _realistic variants run the same staging with a photoreal treatment. It is not "better"; it is a different surface. Choose it when the shot has to sit next to live footage, and skip it when the piece is deliberately stylised — grain and skin texture fight a graphic look.
What the presets do and do not control
They control how the scene is staged, not resolution, duration, or aspect. Those come from the tier: 480P, 768P, 1080P or 1440P, one to fifteen seconds, 16:9 or 9:16.
Three rules that will reject your request if you break them:
- Presets only exist on the multimode variant. A standard-lane job with a preset attached is invalid.
- Multimode supports text-to-video and image-to-video. Text-to-video needs no reference images; image-to-video requires at least one image.
- 1440P implies multimode, and the preset list is the one above — a made-up value is rejected rather than ignored.
A choosing routine
- Is there contact, chase, fall or impact in the shot? Action staging, otherwise drama staging.
- Is the shot meant to cut against photographed actors? Live action, otherwise the animation pair.
- Does the piece need surface texture that reads as film? Add _realistic.
Then write the prompt to match the register. A drama prompt describes what the face does and where the camera is; an action prompt describes the geometry of motion — where the body starts, what it hits, where it ends. Prompts written for one register read as vague instructions to the other.
A note on seeds
The multimode lane accepts an optional seed. That is useful: re-running the same prompt and preset with a different seed helps you compare the results. Hold the seed, change one word; then hold the words, change the seed. Pinning the seed gives the comparison a more controlled starting point; it is not required to submit a job.
FAQ
What do the two staging registers mean? Drama is the dialogue-and-emotion register: faces, blocking, held camera. Action is the choreography register: contact, chase, impact.
Do presets change resolution, duration or aspect ratio? No. Those come from the tier. The preset only changes how the scene is staged.
Can I attach a preset to a prompt-only job? Yes, on multimode text-to-video. Image-to-video still requires at least one reference image.
Where this runs
The six presets are selectable in the H3 generator when the generation style is set to multimode at 480P, 768P, 1080P or 1440P — https://callirra.com/minimax-h3 lists all six tiers and their rates. You do not need to attach references to select a preset for text-to-video.
I work on Callirra, so the preset names and the restrictions above come from the table the product enforces, not from a third-party write-up. If you have run the same prompt across several presets and want to publish the comparison, a measured side-by-side would be more useful than my description — link it in the comments and I will add it.