FL MiniMax Music 3

MiniMax Music 3 nodes for ComfyUI with verified model downloads, full audio VAE loading, MOSS-Music dataset preprocessing, and resumable LoRA training through a pinned SimpleTuner backend.

Features

• Music 3 Loader - Downloads, verifies, and loads the MODEL, CLIP, and decoder VAE with live per-file progress
• Full Audio VAE - Loads the SimpleTuner DAV encoder/decoder for waveform encoding and decoding
• Intelligent Dataset Preprocessing - Uses MOSS-Music to caption, transcribe, analyze, and segment local music libraries
• Large Library Support - Reads an absolute source directory at execution time without uploading or copying the source collection
• LoRA Training - Runs the pinned Music 3 SimpleTuner integration in an isolated environment
• Live Training Dashboard - On-node step, loss, learning-rate, progress, loss-chart, and validation-audio monitoring
• Durable Runs - Preserves logs, metrics, checkpoints, validation samples, interruption state, and resume metadata
• ComfyUI LoRA Export - Copies completed adapters into the standard ComfyUI LoRA folder

Nodes

Installation

ComfyUI Manager

Search for FL MiniMax Music 3 and install it, or open the Comfy Registry listing.

Manual

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI-FL-MiniMaxMusic3.git
cd ComfyUI-FL-MiniMaxMusic3
pip install -r requirements.txt

Restart ComfyUI after installation.

Quick Start

Music 3 Inference

1. Add FL MiniMax Music 3 Loader
2. Choose the model precision, text-encoder precision, CLIP device, and download policy
3. Connect its MODEL, CLIP, and VAE outputs to the standard MiniMax Music 3 workflow
4. Queue the workflow; missing artifacts download only when the node executes
5. Monitor verification, download, and loading progress directly on the node

Dataset Preprocessing

1. Put .wav, .flac, .mp3, .ogg, .m4a, or .aac files in a local directory
2. Add FL MiniMax Music 3 Dataset Preprocessor
3. Paste the directory's absolute path into source_folder
4. Set backend_policy and model_policy to their install/download options for the first run
5. Queue the node and monitor MOSS installation, model download, analysis, and segmentation in its dashboard
6. Connect the resulting dataset directly to the training nodes or review the generated sidecars first

The source directory is scanned recursively at execution time. Files are read in place and are not uploaded through the browser.

LoRA Training

1. Prepare a dataset manually or with FL MiniMax Music 3 Dataset Preprocessor
2. Connect Dataset -> Train Config -> LoRA Trainer
3. Optionally connect Validation Config for checkpoint audio samples
4. Set backend_policy to install_pinned_if_missing for the first run
5. Queue the workflow and monitor the on-node training dashboard
6. Use FL MiniMax Music 3 Training Run to resume an interrupted run from its latest full checkpoint

Completed adapters are saved to:

ComfyUI/models/loras/MiniMaxMusic3/<output-name>/<output-name>.safetensors

Load them with ComfyUI's standard Load LoRA (Model Only) node between the Music 3 loader and sampler.

Dataset Format

Trainer-ready datasets contain one audio file and two text sidecars per segment:

dataset/
  track_001.wav
  track_001.txt
  track_001.lyrics
  track_001.music3.json

• .txt contains the musical caption
• .lyrics contains lyrics and structure tags such as [verse] and [chorus]
• .music3.json contains provenance, source hashes, analysis, and review state for generated segments

The preprocessor writes lossless PCM WAV segments so Windows training does not depend on TorchCodec audio decoding.

Models and Backends

Inference models are stored in the normal ComfyUI model directories. MOSS-Music is checksum-verified under:

ComfyUI/models/moss_music/MOSS-Music-8B-Instruct/

Preprocessing and training dependencies are installed into separate managed environments under:

ComfyUI/user/fl_minimax_music3/backends/

They are not installed into ComfyUI's Python environment. Model and backend downloads begin only after an explicit queued node execution or dashboard action.

Runs and Recovery

Preprocessing and training runs are persisted under:

ComfyUI/user/fl_minimax_music3/

Interrupting ComfyUI requests a graceful worker exit and preserves the latest usable checkpoint. Resume validation prevents accidental changes to dataset identity, LoRA shape, precision, or output name.

Requirements

• A current ComfyUI build with MiniMax Music 3 support
• Python 3.10+ for the node pack; the pinned training backend currently requires Python 3.12 or 3.13
• FFmpeg and FFprobe available on PATH
• NVIDIA CUDA GPU with BF16 support for the initial training backend
• At least 20 GiB free for training-backend installation, plus model and checkpoint storage
• Approximately 16.9 GiB for the optional MOSS-Music model

The first training release targets one NVIDIA GPU. Multi-GPU, LyCORIS, full-rank training, AnyFlow, TwinFlow, CREPA, and LayerSync are not currently exposed.

Example Workflow

Start with example_workflows/MiniMax Music 3 LoRA Training.json. Its deliberately short training and validation settings are intended for installation checks; increase them for real training.

Development Tests

From the ComfyUI directory:

D:\ComfyUI\venv\Scripts\python.exe -m unittest discover -s custom_nodes\ComfyUI-FL-MiniMaxMusic3\tests -q
node --test custom_nodes\ComfyUI-FL-MiniMaxMusic3\tests\*.mjs

License

Apache-2.0. The optional training backend installs and invokes unmodified SimpleTuner as a separate AGPL-3.0-or-later program. MiniMax Music 3, the encoder, and MOSS-Music artifacts retain their respective model licenses. See NOTICE for details.