MiniMax Music Generate

The video nodes in this pack ride on H3's video head. MiniMaxMusicGenerate is the standalone odd one out: it drives the model's generate_music head directly, so you can generate a music track from the same checkpoint that makes your video - text prompt in, raw waveform out, no Suno subscription and no API call. The audio-generation essay in the KB calls music the "bolted-on" layer of the ComfyUI ecosystem, and this node is exactly that pattern: a real model doing real work, wrapped in a bespoke node that depends on the model exposing a nonstandard method.

Which brings us to the single most important thing about this node: it only works if your loaded checkpoint actually has a generate_music head. The source checks for it and raises a clear error - "Loaded checkpoint has no generate_music head - load the joint MiniMax H3 audio+video checkpoint, not a video-only variant." That's the good kind of failure, because it tells you exactly what to fix. If you loaded a video-only H3 variant, this node won't produce silence and pretend; it stops. The flip side is the pack's other placeholder gotcha, inverted: for video, a missing audio head gives you silent audio; here, a missing music head gives you a hard error. Either way, check your checkpoint.

How it works. It grabs the model's audio_sample_rate (default 44100 if absent), builds a CPU/torch generator seeded from your seed, and calls generate_music(prompt, lyrics, style_audio, duration_seconds, cfg, steps, generator). If you supply style_audio (an AUDIO), it moves the waveform to the device and passes it as reference conditioning - that's the "make it sound like this" input. lyrics is optional text conditioning for sung content. The output is an AUDIO dict (waveform + sample_rate) that plugs into MiniMaxMusicPostProcess or into MiniMaxH3VideoExport's audio input for a scored clip.

Inputs that matter:

• prompt (default "upbeat synthwave, driving bassline") - the musical description. This is where the actual direction lives; the model is a music generator, so think genre, tempo, instrumentation.
• duration_seconds (1–300, default 30) - H3 does 4–15s clips for video; the music head stretches further. Keep in mind VRAM and time scale with duration.
• cfg (0–20, default 4) and steps (1–200, default 50) - music sampling knobs. Lower cfg is looser/more creative, higher is more literal to the prompt.
• seed - as ever, the thing that makes a good take reproducible.
• lyrics (optional) and style_audio (optional) - the two extra conditioning paths.

Output. One AUDIO. No latent, no video - wire it to MiniMaxMusicPostProcess for trimming/looping/normalizing, or straight into the export node if you just want a track.

Install. Same as the whole pack:

cd ComfyUI/custom_nodes
git clone https://github.com/TensorVizion/Minimax-H3-Video-Music-Nodes

or via ComfyUI Manager, then restart. Deps: torch, torchaudio, numpy, Pillow. No ffmpeg required here.

Where it bites. Beyond the checkpoint-head requirement, the biggest practical trap is treating it like the API: local H3 is a 33B model, and generating a 300-second track is not a consumer-card activity yet - there's no verified VRAM floor for H3 at all, and the weights (~42.5GB) are the real gate. The license applies here too: the H3 Community License bars running the local weights (and using their outputs) in the US, EU, UK, and Korea, so if that's you, the hosted MiniMax Music-01 API is the sanctioned path. And as a freshly-shipped single-commit pack, this node is a thin scaffold against a model API that the author flags as "check against your build" - test with a short, cheap prompt before you commit to a long one.