ComfyUI Custom Node for Music Flamingo

Custom node for ComfyUI that allows you to analyze songs using NVIDIA's Music Flamingo 2601.

Installation

Clone this repo into custom_nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/tavy315/ComfyUI-Music-Flamingo-Analyzer

Install the node dependencies in the same Python environment ComfyUI uses:
```
python -m pip install -r requirements.txt
```

The node defaults to nvidia/music-flamingo-2601-hf. The first run downloads the model through Hugging Face, so it can take a while.

Nodes

Music Flamingo Analyzer: optimized MVP node for prompt reuse. It defaults to 4-bit NF4 loading with bitsandbytes, nested/double quantization, automatic bf16/fp16 compute dtype selection, trims AUDIO input to a 30-second mono window at the model's native 16 kHz rate, and returns description.

After the first successful model download, enable local_files_only to stop Hugging Face from making remote metadata checks on each fresh ComfyUI process. Keep unload_after_run=false to reuse the in-memory model across repeated runs in the same ComfyUI session.

Prefer quantization=4bit for VRAM-constrained workflows. If you intentionally use quantization=8bit and Transformers reports that modules were dispatched to CPU or disk, enable cpu_offload; it is slower but avoids the 8-bit loader failure.

Use description with a text display or prompt-combiner node, then append your generation suffix for Suno or another music model.

Default analyzer prompts avoid lyric summaries and lyric-theme interpretation.

Example Workflows

workflow/example.json