music_generation

Generating music with MiniMax-Music3

MiniMax-Music3 writes and sings a song from two texts: a caption describing how the music should sound, and lyrics with section tags such as [verse] and [chorus]. This example renders one, either in your own process or through a running MAX server.

The audio generation guide covers the endpoint and the prompt format in more depth; this example is the script that goes with it.

The interesting case is a full-length song. The model denoises 8 second windows that overlap by half, blends them where they meet and crops the result, so a three-minute song is dozens of windows joined at dozens of seams rather than one long generation. The bundled song is 2 minutes 45 seconds, and --check-seams measures those joins.

Requirements

A MAX-compatible GPU. The checkpoint is roughly 28 GiB of weights across its autoregressive, diffusion and vocoder stages, which is more than a 22 GiB card holds at once — the model loads and frees each stage in turn, so a card that size is enough, and the weights download on first use.

Expect the render to take several times longer than the song. On an A10G a 20 second excerpt takes about 5.7x realtime and a 60 second clip about 4.7x, so the bundled 2:45 song is roughly a quarter of an hour. The first render on a machine pays for compiling the graphs on top of that, which took another ten minutes here; later runs replay it from the compilation cache.

Rendering a song

In your own process, with the max package installed:

python generate_song.py --out song.wav

Or through a server, which is worth starting if you plan to render more than one song, since it keeps its compiled graphs between requests:

max serve --model-path MiniMaxAI/MiniMax-Music3
python generate_song.py --out song.wav --server http://localhost:8000

Both paths send the same request. In process it goes through the audio generation pipeline directly; over HTTP it goes to /v1/audio/speech, whose OpenAI schema puts the lyrics in input and the caption in instructions. A server started this way also answers /v1/responses if you set MAX_SERVE_API_TYPES='["openai","responses"]', which returns a URL for the audio rather than the bytes.

A server holds its memory between requests — around 18 GiB on a 22 GiB card — so give it the GPU to itself. A second server, or an in-process render alongside it, fails to allocate.

While you are working on a song, render an excerpt rather than the whole thing:

python generate_song.py --out excerpt.wav --duration 30

--duration sets the length in seconds, --steps the denoising steps per window (30 by default; fewer is faster and rougher), and --seed fixes the sampling so that a render repeats exactly.

Writing your own song

A song is a JSON file. Lyrics are a list of lines so that the file stays readable:

{
  "caption": "Global Metadata: dream pop, 92 BPM, A minor, ... Vocal Details: female lead, airy breathy timbre, ... Arrangement: shimmering electric guitar and analog pad, ...",
  "lyrics": ["[intro]", "", "[verse]", "Headlights paint the empty road"],
  "duration": 165.0,
  "seed": 1235
}

Pass it with --song my_song.json.

The model card asks for the caption in three sections — Global Metadata, Vocal Details, Arrangement — and reads the section tags in the lyrics. Both are worth following: at full length the arrangement is where you say how the song should develop, and a song that says nothing about its development has no reason to develop. Tempo in the caption is a hint the model may not take literally.

Checking the seams

A join that went wrong sounds like a click or a lurch in level. Both are cheap to look for, because the offsets follow from the model's constants rather than from the audio:

python generate_song.py --out song.wav --check-seams

Neither measurement means anything on its own — music is full of transients — so each seam is scored against thousands of arbitrary offsets in the same render, and the verdict comes from how many seams clear a high quantile against how many chance predicts. To check a WAV you already have, without paying for the render again:

python generate_song.py --analyze song.wav --duration 165

Running with Bazel

From a checkout of the repository:

./bazelw run //max/examples/music_generation:generate_song -- --out song.wav