The first time I used a style preset on MiniMax H3, I got a video of two people standing very close to each other, breathing heavily, doing nothing.
The prompt was a fight. The preset was a drama preset. The preset won.
That is the thing about these six values: they are not colour filters. They change how the model stages a scene — where bodies go, where the camera sits, whether a shot is built around a line of dialogue or around an impact. Choose the wrong axis and you get a technically clean video of the wrong genre.
I work on Callirra, which is where I ran into this. The presets are selectable in the generator and set through one API field, and they are worth understanding before you write a prompt rather than after.
The six values
| Value | Staging | Treatment |
|---|---|---|
live_action_drama |
dialogue and emotion | live action |
live_action_drama_realistic |
dialogue and emotion | live action + photoreal pass |
live_action_action |
contact, chase, impact | live action |
live_action_action_realistic |
contact, chase, impact | live action + photoreal pass |
anime_drama |
dialogue and emotion | animation |
anime_action |
contact, chase, impact | animation |
Two axes: staging and treatment. Read them separately and you do not need to memorise names.
"She reads the letter and looks up" is drama staging in either treatment. "He vaults the railing and lands running" is action staging in either treatment.
Setting it through the API
The presets live on the multimode path, which accepts prompt-only text-to-video as well as image-to-video, and covers 480P, 768P, 1080P and 1440P:
curl https://api.callirra.com/v1/videos \
-H "Authorization: Bearer $CALLIRRA_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "minimax-h3",
"prompt": "she reads the letter, folds it, and looks up at the door",
"resolution": "1080P",
"duration_seconds": 8,
"aspect_ratio": "16:9",
"minimax_h3_variant": "multimode",
"minimax_h3_style_preset": "live_action_drama"
}'
An unknown value is rejected rather than ignored — the behaviour you want, because a silently coerced preset is a shot you cannot explain later.
What each staging register actually does
Drama staging is the dialogue-and-emotion register. Characters stand, sit, listen, turn. The camera holds or drifts. The drama is in faces and blocking. If your prompt is about a decision, a reveal, a pause, this is the register.
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 contains a verb that changes someone's position in space, this is the register.
Feed a dialogue beat to an action preset and you get unmotivated movement: the model stages something because staging is what that preset is tuned to produce. Feed a fight to a drama preset and you get the two-people-breathing-intensely shot I opened with.
The photoreal pass is a grading decision
The _realistic variants run the same staging with a photoreal treatment. They are not "better". They are a different surface.
Use them when the shot has to intercut with live footage — grain and skin texture are what make the cut invisible. Skip them when the piece is deliberately stylised, because that same texture fights a graphic look.
Where this goes wrong
- Choosing the preset after writing the prompt. It is a staging decision; decide it from the beat, not from the render you did not like.
- Expecting consistency. Presets shape staging, not identity. Keeping a specific face or product across shots is the reference system's job.
- Using them on 2K. The presets are a multimode feature; 2K is a different path with different capabilities.
- Combining presets with paired frames. Different control surfaces. Do not assume one holds the other.
A short decision path
- Write the beat as one sentence.
- Conversation → drama. Movement → action.
- Intercutting with real footage → realistic. Graphic or stylised → plain.
- If the answer changes between takes, the problem is the script, not the preset.
FAQ
Do presets change the price? No. Price follows the resolution tier, not the preset.
Can I use them with text-only prompts? Yes on the multimode path, which accepts prompt-only jobs.
What happens with an unknown value? It is rejected rather than coerced to a default.
Do the API values match the UI labels? The UI shows readable labels such as Live action · Drama; the API takes the identifier in the table above.
The presets and their labels are in Callirra's H3 generator. If you have a beat that a preset stages in a way this post says it should not, that counter-example is worth more than my description.
Top comments (5)
The staging vs treatment split is the useful part — read that way the six values stop being names to memorise.
The detail about the preset winning over the prompt is the real lesson here — it changes how you write the prompt, not just which preset you pick. Curious whether you've found any prompt patterns that consistently override a preset's staging.
The 'not colour filters' framing is the part that stuck with me — it changes how you write the prompt, not just which preset you pick.
The two-axes framing is what made this click for me.
I'd like to see a comparison where the same prompt is run through drama and action presets side by side — that would make the staging difference obvious.