MiniMax Music 3 Generation Settings

Reproducibility in a music-generation graph has a chicken-and-egg problem: you want the text stage and the audio sampler stage to each get their own seed, but if you hand-type them separately, the two seeds drift apart and "same song again" becomes "slightly different song every time." MiniMax Music 3 Generation Settings solves it the boring, correct way - you give it one generation_seed, and it derives both stage seeds from it. Change one number and the whole run changes deterministically; keep it and everything lines up.

The derivation is dead simple: text_seed = generation_seed, and ksampler_seed = text_seed + ksampler_seed_offset (both modulo the seed range). Offset 0 keeps them aligned, which is the default and the right call for most people; a nonzero offset is for when you want to vary the audio sampler while keeping the same primary seed reference. This is the reproducibility anchor of the whole pack - the README calls generation_seed the seed you can also reuse for artwork generation.

The inputs

• generation_seed - the primary song seed. Force-input, so wire it from the parser or prompt source.
• max_duration (default 300 s, cap 360) - an upper bound, not a promise. The model can still end early if the Lyrics/section structure encourages a shorter track, which is why the prompt library harps on long instrumental section maps.
• text_cfg_scale (default 1.7) and text_top_k (default 50) - guidance and sampling breadth for the MiniMax text/autoregressive stage.
• ksampler_seed_offset, ksampler_steps (default 40), ksampler_cfg (default 1.7), denoise (default 1) - the diffusion/audio sampler stage.

Outputs

Everything the MiniMax node and the metadata chain needs, as typed outputs: max_duration, text_seed, text_cfg_scale, text_top_k, ksampler_seed, ksampler_steps, ksampler_cfg, denoise. Wire the text values into the MiniMax text stage, the sampler values into the audio sampler, and all of them into MiniMax Song Metadata so the sidecar records exactly what ran. (There's a separate KSampler + Config Output node for the FLUX cover branch - this node is specifically the MiniMax Music 3 settings hub.)

Installing it

Pack install via ComfyUI Manager (search "MiniMax Music Production Toolkit") or:

cd ComfyUI/custom_nodes
git clone https://github.com/jplenio/ComfyUI-MiniMax-Music-Production-Toolkit.git
cd ComfyUI-MiniMax-Music-Production-Toolkit
python -m pip install -r requirements.txt

Dependencies: scipy, soundfile, imageio-ffmpeg, mutagen, Pillow. Restart ComfyUI and hard-refresh once.

Gotchas

The 360-second max_duration cap is the one hard limit that surprises people - MiniMax Music 3 won't be pushed past it from here, so if you genuinely need a six-minute track, this node isn't the place to ask for it. And note max_duration is a float on input but the audio sampler expects what it expects; the node passes it through as a FLOAT, so check your downstream node's accepted type if you hit a validation error. Otherwise there's not much to break here - which is the point. It's a deterministic settings hub, and it behaves like one.