boo-moss-audio

ComfyUI custom nodes for OpenMOSS's MOSS-Audio audio-understanding models — speech transcription, music/mood analysis, and general audio captioning/QA in one model.

Nodes

• BOO MOSS-Audio Loader — picks one of the four released checkpoints (MOSS-Audio-4B-Instruct, MOSS-Audio-4B-Thinking, MOSS-Audio-8B-Instruct, MOSS-Audio-8B-Thinking) and downloads it from Hugging Face on first use to ComfyUI/models/moss-audio/<model-name>. Outputs a BOO_MOSS_AUDIO_MODEL.
• BOO MOSS-Audio Generate — takes a BOO_MOSS_AUDIO_MODEL, a native ComfyUI AUDIO input, and a text prompt, and returns the model's text response as a STRING. The default prompt asks for explicitly labeled LYRICS:/STYLE: sections — plain requests like "transcribe the lyrics, then describe the mood" reliably get the model to transcribe the lyrics but drop the mood/style description entirely once it runs out of words to transcribe; forcing two required, labeled sections fixes that (verified empirically across both greedy and sampled decoding). Thinking-variant <think>...</think> reasoning blocks are stripped by default (strip_thinking).

Wire BOO MOSS-Audio Generate's STRING output into any downstream prompt-enhancement node (e.g. boo-textgen) to feed a transcript/mood description into further prompt generation.

Example workflow

example_workflows/basic_caption.json — LoadAudio → BOO MOSS-Audio Loader + BOO MOSS-Audio Generate → Preview as Text. Load it in ComfyUI (Workflow → Open), pick an audio file in the LoadAudio node, and queue the prompt.

Requirements

Runs inside a ComfyUI checkout, same as any custom node package. The only extra dependency is huggingface_hub, used to download checkpoints — see pyproject.toml/requirements.txt. torch/torchaudio/transformers already bundled with ComfyUI are used as-is.

Vendored MOSS-Audio code

vendor/moss_audio/ contains four small files copied from the MOSS-Audio repository (modeling_moss_audio.py, processing_moss_audio.py, configuration_moss_audio.py, audio_io.py) rather than a pip install of the upstream package. Two reasons:

1. MOSS-Audio's Hugging Face checkpoints only declare trust_remote_code support for AutoConfig/AutoProcessor in config.json, not AutoModel — there's no modeling code bundled in the model repo itself, so AutoModel.from_pretrained(..., trust_remote_code=True) fails with "Unrecognized configuration class". The upstream repo's own infer.py works around this by importing MossAudioModel directly from its src/ package; these vendored files let this ComfyUI package do the same.
2. pip installing the upstream moss-audio package pulls in gradio and streamlit (used only by its own demo app.py) as unconditional base dependencies, which force-upgrades huggingface_hub to a version incompatible with the transformers version ComfyUI itself depends on — this actually broke ComfyUI's own model loading when tried. Vendoring just the inference-path files avoids that dependency tree entirely.

See vendor/moss_audio/NOTICE for provenance, source commit, license basis (Apache-2.0), and the one line changed from upstream (an import path fix).

Tests

source /path/to/ComfyUI/.venv/bin/activate
pytest

GPU integration tests

tests_gpu/ holds an opt-in suite that downloads a real checkpoint and exercises a full load → offload → reload → generate cycle against an actual CUDA GPU, verifying the model survives being moved between devices by ComfyUI's ModelPatcher. It's excluded from the default pytest run (not in pyproject.toml's testpaths) since it needs a GPU, network access, and downloads a multi-gigabyte checkpoint. Run it explicitly:

pytest tests_gpu/