ComfyUI-MOSS-SFX
Sound effects from an English prompt in ComfyUI with MOSS-SoundEffect v2.0. Runs locally, works on 8 GB cards.
ComfyUI-MOSS-SFX (version 1.0)
Small ComfyUI nodes that make a sound effect from an English prompt with MOSS-SoundEffect v2.0 by OpenMOSS. Runs locally, works on 8 GB cards.
The model code comes from OpenMOSS's moss_soundeffect_v2 package (Apache-2.0). These nodes only call that package and convert the result to ComfyUI AUDIO.
A Korean guide is available in README_KO.txt.
Install
Run every command in a cmd window opened in the ComfyUI_windows_portable folder (the one that contains run_nvidia_gpu.bat), one line at a time. Git is required for the install (check with git --version).
- Get the node into
ComfyUI\custom_nodes, either with git:
git clone https://github.com/Linchwe/ComfyUI-MOSS-SFX ComfyUI\custom_nodes\ComfyUI-MOSS-SFX
or by downloading the ZIP (Code > Download ZIP) and extracting it so that the result is ComfyUI\custom_nodes\ComfyUI-MOSS-SFX\nodes.py.
- Record the current state of your Python packages. Skip this line if
pip_before.txtalready exists.
python_embeded\python.exe -m pip list --format=freeze > pip_before.txt
Keep this file. The -c pip_before.txt at the end of the install lines tells pip not to change versions of packages you already have. You can also roll back with python_embeded\python.exe -m pip install -r pip_before.txt.
- Install the required packages. Run the 4 lines in order, one at a time, and keep
-c pip_before.txt.
python_embeded\python.exe -m pip install --no-deps git+https://github.com/OpenMOSS/MOSS-TTS.git#subdirectory=moss_soundeffect_v2 -c pip_before.txt
python_embeded\python.exe -m pip install ftfy einops diffusers soundfile imageio regex -c pip_before.txt
python_embeded\python.exe -m pip install --no-deps descript-audiotools==0.7.2 -c pip_before.txt
python_embeded\python.exe -m pip install argbind pyloudnorm importlib-resources julius ffmpy ipython rich matplotlib librosa pystoi torch_stoi flatten-dict markdown2 randomname tensorboard -c pip_before.txt
Red version-mismatch warnings during install are expected as long as the last line says Successfully installed. Verify with python_embeded\python.exe -m pip show moss-soundeffect-v2.
- Download the model (about 11 GB).
python_embeded\python.exe -c "import sys, huggingface_hub as h; h.snapshot_download(sys.argv[1], local_dir=sys.argv[2])" OpenMOSS-Team/MOSS-SoundEffect-v2.0 ComfyUI\models\moss_soundeffect_v2\MOSS-SoundEffect-v2.0
ComfyUI\models\moss_soundeffect_v2\MOSS-SoundEffect-v2.0\model_index.json must exist.
- Fully restart ComfyUI.
Nodes
Double-click the canvas and search for "MOSS SFX".
MOSS SFX Loader -> MOSS SFX Generate -> Save Audio (built-in)
- memory_mode: use
low_vramfor 8 GB cards.standardalso works with 12 GB or more. If unsure, uselow_vram. - dit_dtype:
bfloat16(default) for RTX 30 series or newer,float16for RTX 20 series. Both use the same memory.float32uses about twice the memory; avoid it on 8 GB cards. - keep_loaded: if off, the model is unloaded from VRAM after each generation (slower next run).
- unload_other_models: on by default. Unloads models ComfyUI keeps in VRAM (e.g. an image model used earlier) before generating. Keep it on for 8 GB cards.
- Prompts: English or Chinese only.
Notes
- The first generation takes about 3 minutes while the model loads. Later runs reuse the loaded model.
- Peak VRAM in
low_vrammode was about 4-5 GB (measured on an RTX 3060 12GB). The 8 GB / RTX 20 series settings above are not yet fully verified. - If the console shows
[MOSS-SFX] This GPU has no bfloat16 hardware, set dit_dtype to float16. - 16 GB or more of system RAM is recommended.
License
These nodes are MIT licensed (see LICENSE). The moss_soundeffect_v2 package and the MOSS-SoundEffect v2.0 model are Apache-2.0 (see NOTICE).