MiniMax Music 3 Audio Generate ⚡

Music is the corner of local generation where the API (Suno, in practice) has historically owned "reliable." This node is the counterpoint: run MiniMax Music 3's full AR + flow DiT + DAV pipeline on your own GPU, caption plus lyrics in, a 44.1 kHz stereo track out, with everything offline after the models are downloaded. It's not "Suno at home with a worse model" - it's a genuinely different trade: you control the seed, the duration, the sampling knobs, and you don't upload your lyrics anywhere.

What it does

Feed it the model / clip / vae from MiniMax Music 3 Models Loader, plus two text fields, and it runs the whole chain:

• caption - global metadata: genre, BPM, key, mood, production, vocal character, instruments, arrangement. The default is a well-structured example (dreamy synth-pop, 112 BPM) - copy its shape and you'll get better results than prose.
• lyrics - section-tagged lines. Tags it understands: [intro], [verse], [pre-chorus], [chorus], [bridge], [instrumental], [drop], [outro]. Leave it empty for an instrumental.

The knobs that matter

• max_duration - seconds, up to 360 (6 minutes). Read the tooltip twice: it's an upper bound. The autoregressive model can decide the song is musically complete and end it earlier. That's a feature - songs that end on a cadence beat looping forever - but it means 120 won't guarantee you 120.
• seed - reproducibility. Same caption, same seed, same song.
• steps / guidance_scale - the DiT sampling knobs (30 steps, 1.7 CFG defaults).
• ar_cfg_scale / ar_top_k - the autoregressive conditioner's knobs (1.5 / 50). Lower top-k tightens the "vocabulary" of the plan; it's the closest thing here to a "how weird" dial.
• decode_mode - auto (default), tiled, or dense. DAV decoding flips to tiled mode automatically for long outputs to keep memory in check; tile_size (1536) and tile_overlap (64) only matter if you're forcing tiling by hand.

One output

audio (AUDIO) - 44.1 kHz stereo, straight into Save Audio, a video mux, or any AUDIO input in ComfyUI. This is standard comfy AUDIO, so it composes.

Installing it

Part of the ComfyUI-GGUF-Loader pack under 🤖 CCTech/MiniMax Music. ComfyUI Manager → search "ComfyUI-GGUF-Loader" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ChrisColeTech/ComfyUI-GGUF-Loader
cd ComfyUI-GGUF-Loader
pip install -r requirements.txt

Then download Comfy-Org/MiniMax-Music-3 and place its split files per the loader's tooltips (DiT → models/diffusion_models, pruned AR encoder → models/text_encoders, DAV → models/vae).

Common issues

Where people get burned: expecting max_duration to be exact, and treating lyrics like freeform text instead of tagged sections - untagged lyrics still work but you lose the structural control the tags buy you. If you want an instrumental, empty lyrics, not a blank-tagged mess. And if the sampler seems to stall on a long max_duration, check whether it's in tiled decode - the automatic switch is meant to handle exactly that, so leave decode_mode on auto unless you have a reason. Run tools/smoke_minimax_music3.py from the pack to validate the three load paths before blaming the sampler.