Everything MiniMax H3 can do

MiniMax H3 is MiniMax's video model. Every clip comes with its own sound, made from the same prompt as the picture: effects that land on the frame they belong to, music, and voices that say the lines you quote. A step can read up to nine pictures, three clips and three sound files, twelve in all, to keep a face or a look, borrow a camera move, or edit a clip you already have. Clips run up to 15 seconds and up to 1440p, in four versions: Standard, Fast, Max and Max Turbo.

Every clip below was made by the step under it. Change a placeholder and press Gen to make your own.

Many beats, one take

A long clip needs something to happen in every part of it. Write the beats with their seconds ("0 to 3 seconds: …") and name the sounds in the same order. Duration goes up to 15 seconds; one action stretched that far tends to stall or repeat.

Edit a clip

Give it a clip and an instruction, and only what you name changes: the cat sleeps on, the camera pushes in as before. Say what stays as plainly as what changes. Change the sound with the picture: after dark, crickets take over from the cicadas.

Each edit's Duration follows its input, so the edit runs as long as the clip it changes.

First and last frames

Give it a first and a last frame and it makes the shot between them. Make the last frame an edit of the first, changing one thing, so everything else lines up, and use the prompt for what happens on the way. A step reads frames or references, not both. These stills come from FLUX 3 Image too.

With one frame, the clip starts there and takes the photo's shape: a tall photo makes a tall clip. Ask for small motion the photo already suggests, a blink, a breath, a breeze; a big action pulls the face away from the photo. Drop in your own.

Borrow a camera move

Give a step a clip and it can lift the camera move out of it and use it on something else. Describe the new subject, not the move: the prompt only says to follow the clip's camera path, so the half circle comes from the clip alone. A smooth move around one plain subject carries over best.

A clip a step reads is billed by the second, on top of the clip it makes.

Give each picture a job

A step can read up to nine pictures. In the prompt, give each one its job: here one is the woman and the other her dog. Then name what must carry over, her face, hair and clothes and the dog's markings and collar, and let the words make the rest: the beach, the run, the sound. Both pictures come from FLUX 3 Image steps on this page.

A picture can also lend its style alone. Say so in the prompt and the clip takes its materials, colors and light into a scene of its own. This step runs on Fast, which reads pictures and clips as Standard does and makes 480p clips. Drop in any picture as the style.

Quote the line

Put the spoken line in quotes and the clip says it, the mouth moving with the words. Keep it to one breath, say how it's said (cheerfully over the market, firmly over the wind), and write what the speaker does next, so the clip doesn't end on a frozen face.

Write the sound

Every clip comes back with sound, and there is no setting for it: the prompt directs it. End the prompt with a Sound: line that names each sound in the order it happens and ties it to its moment: "the whistle's shriek as the engine bursts from the tunnel", "a thunderous splash exactly as the whale lands". Ask for each one loud and close; a sound asked for quietly may not come through.

Keep the picture and change only the music, and the same shot gets a new score. Describe the music by its instruments, tempo and mood, never by a song or an artist, and say "no voices" when the picture should lead. These three run on Max Turbo, the version that reads no references, at half the price of Standard at 768p.

To build on any step, press Mix to copy its recipe into your own page.