For about a week, every quiet scene I generated had somebody shifting their weight in it.
Not ruining the shot — just animating it, the way a nervous actor does. I rewrote the prompt four times. I removed the verb. I asked for stillness in three different ways. It kept happening, and I kept blaming my wording.
It was the preset. The job was running action staging, and action staging stages movement, because that is the job it was given.
The short version. H3 ships six scene presets across two axes — drama or action staging, live action or animation — available only on the multi-mode path at 480P, 768P, 1080P and 1440P. They change staging, not resolution or price.
What do drama staging and action staging actually mean?
The product labels two staging registers: drama and action. In Chinese production vocabulary they are wenxi (drama) and wuxi (action) — not genres, but descriptions of what a single scene is doing. A whole film has both, and this register is the axis the six presets are built on.
Once I saw it that way, the six presets stopped looking like flavours and started looking like a routing decision.
What are the six scene presets?
There are two axes: staging (drama vs action) and treatment (live action vs animation). The photoreal variants sit on top of live action as a grading choice, not a quality tier.
- Live action · Drama — faces, blocking, a camera that holds or drifts.
- Live action · Realistic drama — the same staging with skin, fabric and light pushed toward footage.
- Live action · Action — contact, chases, stunts.
- Live action · Realistic action — same choreography, photoreal surface.
- Animation · Drama — animated, drama staging, character art that stays consistent.
- Animation · Action — animated, action staging: fast cuts, impact frames, stylised motion.
The photoreal pair are for shots that have to cut against real footage. Grain and skin texture fight a graphic look, which is why I do not use them by default.
The failure is a category error, not a bad prompt
Give an action preset a letter-reading scene and it invents movement to justify itself. Give a drama preset a fight and you get two people standing very close together with a lot of intensity. Both look like model problems. Both are me handing over the wrong script.
My rule now is one question, asked before I write anything: does anything in this shot touch anything else? Contact, chase, fall, impact — yes means action staging. No means drama staging.
Then I write the prompt to match. Drama prompts describe what the face does and where the camera sits. Action prompts describe the geometry: where the body starts, what it contacts, where it ends. A drama prompt handed to an action preset reads as vague instructions to a choreographer.
Which requests get rejected?
Presets only exist on the multi-mode path — 480P, 768P, 1080P and 1440P, up to fifteen seconds. Attach one to a standard job and it is rejected rather than quietly ignored. 1440P implies multi-mode, so it always has them. A made-up preset value is rejected, not defaulted.
What presets do not do
They do not change resolution, duration, ratio or price. They do not fix object fidelity — a hero prop still needs its own reference and a slower shot. They do not hold a conversation that turns into a chase; that is two takes cut together.
And they will not rescue a scene that needs both registers. The moment a fight has dialogue in it, I split the shot.
All six presets and the tiers that carry them are listed on the H3 page.
I work on Callirra, which is how I can see the staging rules rather than infer them from output. What I do not have is one prompt rendered across all six presets; if you run that, it is worth more than this piece, and I will link it.