MiniMax H3 on SceneMixer: two tiers, priced per second

What sets the family apart

Three things. The picture: the top tier renders at 2K — the larger pixel count of the two — and MiniMax H3 Fast renders at 768P for drafts and phone screens. The cut: a segment can hold several shots with real cuts between them inside one generation, so an exchange and the reaction to it can live in a single render; the storyboard’s shot list is the edit. The sound: every generation includes dialogue and ambient sound — there is no silent mode.

The tiers are picked the usual way: the button in the episode editor’s top bar, or Settings → Model Preferences → Video Model. Whatever you set governs what you render next — a segment already made is untouched, and each version remembers the tier behind it.

What each tier costs per second

Two tiers on a single bill, charged per second of rendered picture.

What changes between the two is the resolution and the rate; the segment ceiling and the reference budget belong to the family and are shared by both. A draft on MiniMax H3 Fast can be re-rendered as a keeper on MiniMax H3 without touching anything else.

What you can make with it

Above is a still from The Glass Floor, which this site took from script to composite on Wan 3.0 · 720P. The setting carries the same flow to a MiniMax H3 tier — reference budget permitting.

Where the family earns its place: a short, dense beat — a confrontation that turns, a reveal with its reaction, anything that fits in a single segment and wants the larger picture; and a segment built from a few cuts, since the shots can be cut inside the generation rather than stitched together from separate renders.

There is room for several episodes in a project too: they sit in order, each rendering as its own chain of segments, and any shot inside them can pick its own tier.

The reference budget, in practice

This family keeps its own numbers: up to nine images and three voice clips per generation, with about 14 s of voice in total. A voice clip also needs an image alongside it — a segment that would carry voices with no image has its voice clips dropped, and each speaker falls back to the written voice description.

When a segment would go over a limit, the site leaves out the least important references first, so the usual failure is a voice changing or a prop losing its sheet rather than a refusal. A segment that names a crowd and attaches a sheet for every face will hit the cap; fewer, more deliberate references carry further here. Segments that need the whole cast attached are better kept on a Wan step, where the budget goes to ten images and five voice clips.

How to start

Count the cast before the tier switch: a MiniMax H3 segment will take up to nine images and three voice clips, and no more — a scene that wants a sheet for every face may need trimming first, or it belongs on a Wan step. When the count works, Select Model is the switch, and only the segments rendered afterwards move.

Money-wise the family bills like the rest of the site: the read of the script, the reference sheets beyond each account’s free allowance, and the rendered seconds each carry their own charge. The purpose-by-purpose ranking places both H3 tiers, and the H3-or-Wan comparison is the longer read.

FAQ

Updated October 14, 2026