MiniMax H3

How H3 writes sound

Audio is generated from the prompt, not pulled from a library. Everything before the Sound: clause describes the picture, and everything after it describes the track. List whatever the scene should contain: ambience, individual effects, music, or spoken lines. There is no separate audio parameter and no extra call, so the same prompt that directs the shot also directs its sound.

import { createClient } from '@runware/sdk'

const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY })
await client.connect()

const [result] = await client.run({
  model: 'minimax:h3@0',
  positivePrompt: 'A hand pops a champagne bottle at a bright celebration, the cork flying and pale-gold foam surging over the neck. Sound: a loud festive cork pop exactly as the cork flies, then the fizz of overflowing foam and distant cheering.',
  width: 1440,
  height: 2560,
  duration: 5
})

import asyncio
import os

from runware import Runware


async def main():
    async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client:
        results = await client.run({
            "model": "minimax:h3@0",
            "positivePrompt": "A hand pops a champagne bottle at a bright celebration, the cork flying and pale-gold foam surging over the neck. Sound: a loud festive cork pop exactly as the cork flies, then the fizz of overflowing foam and distant cheering.",
            "width": 1440,
            "height": 2560,
            "duration": 5
        })


asyncio.run(main())

curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "videoInference",
      "taskUUID": "c8d9e0f1-2a3b-4c4d-9e5f-6a7b8c9d0e1f",
      "model": "minimax:h3@0",
      "positivePrompt": "A hand pops a champagne bottle at a bright celebration, the cork flying and pale-gold foam surging over the neck. Sound: a loud festive cork pop exactly as the cork flies, then the fizz of overflowing foam and distant cheering.",
      "width": 1440,
      "height": 2560,
      "duration": 5
    }
  ]'

runware run minimax:h3@0 \
  positivePrompt="A hand pops a champagne bottle at a bright celebration, the cork flying and pale-gold foam surging over the neck. Sound: a loud festive cork pop exactly as the cork flies, then the fizz of overflowing foam and distant cheering." \
  width=1440 \
  height=2560 \
  duration=5

{
  "taskType": "videoInference",
  "taskUUID": "c8d9e0f1-2a3b-4c4d-9e5f-6a7b8c9d0e1f",
  "model": "minimax:h3@0",
  "positivePrompt": "A hand pops a champagne bottle at a bright celebration, the cork flying and pale-gold foam surging over the neck. Sound: a loud festive cork pop exactly as the cork flies, then the fizz of overflowing foam and distant cheering.",
  "width": 1440,
  "height": 2560,
  "duration": 5
}

[
  {
    "taskType": "videoInference",
    "taskUUID": "c8d9e0f1-2a3b-4c4d-9e5f-6a7b8c9d0e1f",
    "videoUUID": "9f0a1b2c-3d4e-4f5a-9b6c-7d8e9f0a1b2c",
    "videoURL": "https://vm.runware.ai/video/os/a14d18/ws/2/vi/9f0a1b2c-3d4e-4f5a-9b6c-7d8e9f0a1b2c.mp4"
  }
]

The Sound: label is a convention, not a keyword, so a plain sentence describing the audio works too. Keeping it as a labelled clause at the end just makes prompts easier to read and edit.

Ambience is the bed of a scene

The simplest audio to direct is the constant layer under everything: the room tone, the weather, the environment. Name where the shot is and what that place sounds like, and H3 lays it down as a continuous bed. Here a rainy cafe carries four ambient elements at once.

Rain, murmur, cup clinks, and low jazz sit together without any of them being an event. Ambience is what you reach for to give a shot a sense of place, and it costs nothing extra to stack several layers in one clause.

Sound effects, synced to the action

This is the single most useful trick in the model. Naming a sound is easy, but naming when it happens is what ties it to the frame. Words like "as", "when", and "then" bind an effect to a moment, so the sound lands with the picture instead of floating over it.

The prompt did not just ask for a pop, it asked for a pop "exactly as the cork flies", then a fizz, then cheering. That ordering is the instruction H3 follows to place each sound. Describe your effects as a small timeline tied to what is on screen, and the model syncs them for you.

Music and mood

Ask for a track and describe it the way you would brief a composer: genre, tempo, mood, and whether it has vocals. Music is what sets the emotional read of a shot, and the same footage carries a different message under a calm piano than under an upbeat electronic beat.

Adding "no voices" kept the bed clean under a product shot, which matters when the picture has to carry a brand. For a mood you can swap, keep the visual prompt fixed and change only the music description, and the same clip re-scores to a new feeling.

Dialogue and lip-sync

Put a spoken line in quotation marks in the prompt and H3 both speaks it and moves the mouth to match. This turns the model into a one-shot talking-head tool: a presenter, a character, or a narrator delivering a line with the picture, no separate voiceover pass.

The line she speaks is the one in the prompt, and her mouth matches it. Keep spoken lines short enough to fit the clip: a sentence or two suits a five-second shot, while a paragraph will feel rushed or run past the end. Describe the delivery too, since "clearly and warmly" shapes the read as much as the words do.

Where it fits

Native audio is H3's clearest advantage, and it is strongest on concrete, physical sound: foley, ambience, and effects tied to a visible action. Music and mood follow a brief well. Dialogue is the most demanding, since a long line, an unclear delivery note, or a clip too short for the words will strain the lip-sync. Keep spoken lines brief, describe the delivery you want, and let the picture and its sound come out of one prompt.

Tips

1. End with a Sound: clause. Describe the picture first, then the audio. It keeps prompts readable and makes the track easy to edit on its own.

2. Tie effects to the action. "A pop as the cork flies" syncs, while "a pop" floats. Timing words are what place a sound on the frame.

3. Stack ambience freely. Rain, murmur, and music can share one bed, and layering a few environmental sounds gives a shot its place.

4. Brief the music by style. Genre, tempo, mood, and instrumentation, plus "no voices" when the picture needs to lead. Describe the sound you want, never name a song or an artist, and swap only the music description to re-score the same clip.

5. Quote dialogue and keep it short. A line in quotation marks is spoken and lip-synced, and a sentence or two fits a five-second clip better than a paragraph.