ComfyUI Custom Node for Music Flamingo

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

Installation

Clone this repo into custom_modules:

cd ComfyUI/custom_nodes
git clone https://github.com/C0untFloyd/comfyui-musicflamingo

Install dependencies:
```
pip install -r requirements.txt
```
(Requires transformers>=5.0.0, torch, torchaudio, bitsandbytes, accelerate)

Quantization & Performance

The Music Flamingo Analysis node includes a quantization input:

none (default): Runs in FP16/BF16 (~15GB VRAM required).
8-bit: Uses BitsAndBytes INT8 quantization (~8GB VRAM).
4-bit: Uses BitsAndBytes NF4 4-bit quantization (~4.5GB VRAM).

Offline Pre-Quantization (Recommended for Instant Startup)

By default, if no local quantized checkpoint is found, the node will quantize the model on-the-fly on first load.

To eliminate first-load delay and avoid storing full-precision models on disk, you can pre-quantize and save the checkpoint once using the included export script:

# Export 8-bit checkpoint
python custom_nodes/comfyui-musicflamingo/export_quantized.py --bits 8

# Or export 4-bit checkpoint
python custom_nodes/comfyui-musicflamingo/export_quantized.py --bits 4

The script saves the quantized weights to: ComfyUI/models/checkpoints/musicflamingo/music-flamingo-8bit/ (or 4bit/).

Once exported, ComfyUI will automatically detect and load the pre-quantized model directly from disk whenever 8-bit or 4-bit is selected in the node!

Export via Google Colab:

If you don't have enough local bandwidth or GPU VRAM to run the export locally, open export_quantized_colab.ipynb in Google Colab (with a free T4 GPU). It will download, quantize, and package the weights into a ZIP or save them directly to your Google Drive / Hugging Face Hub.