Scripted video with sound on MiniMax H3

MiniMax H3 renders a video and its soundtrack in one pass. You can describe a shot in one sentence, but you can also script it by time: each beat of the prompt says what happens, and when. That is how you get a shot that goes somewhere - a door opens, then a cat walks in, then it meows - instead of one moment stretched over five seconds.

The Director turns that kind of script into the shot-by-shot brief H3 was trained to read. This guide shows the pattern, where the switch is, and sets up your first render in about a minute.

1. The prompt pattern

Put each beat in a time range, in seconds, in order:

[0-2s] A red door in a white wall creaks open and morning light spills in.
[2-4s] A grey cat steps through, stops, and looks up at the camera.
[4-5s] It meows once; birds chirp outside.

• One action per beat.“A cat steps through, stops, and looks up” is about the most one beat should carry.
• Give each beat time. Around 1-2 seconds or more. Five beats in five seconds tends to blur.
• Write the sound where it happens.H3 generates audio from the same prompt, so “it meows once” in the last beat puts the meow at the end.
• Cover the whole clip. The ranges should add up to the duration you pick - 5, 7, 10 or 15 seconds.

2. Keep the Director on

The Directorpill on the composer bar (on by default) rewrites your script into the structured, shot-by-shot brief H3 was trained to read - camera moves, cut times, the soundscape - and reads any picture you attach. The brief it wrote is shown on the result; “Edit brief and re-render” lets you change a line of it and render again. A prompt you already wrote in H3's own format is kept as written and only enriched.

• It adds about a minute before the render starts, and costs nothing extra.
• It works on all H3 modes: Text → Video, Image → Video and Reference → Video.

What we have measured: it is best at events in time and directed motion - a lamp switching on halfway, ducks crossing left to right - and can over-widen a shot whose subject is small. The fair test is yours: set a seed under More, render once with the Director off and once with it on, and compare.

3. What the example renders

4. What it costs

The example renders at 480p for 5 seconds, which fits inside the 10 free Takes every new account gets. At 768p - the resolution H3 is trained for - 5 seconds is 20 Takes. Either way the composer shows the exact price next to Generate before you press it, and a render that fails refunds itself.

A good habit: get the script right at 480p, then render the keeper at 768p. If nothing is warmed up when you press Generate, the first render takes a few minutes to start; while it stays warm, the next ones start straight away.

5. Beyond text

• Image → Video: start from your own still and script what happens next. Add a last frame to decide where it ends.
• Reference → Video: give it pictures, clips or a voice, refer to them in the script as <Picture 1>, <Audio 1>, and time their beats the same way.
• Single-moment shots - a slow pan across a landscape - do not need a script; one sentence is fine.