Load MOSS Sound Effect Model

MossSoundEffectLoadModel is the plain loader half of the sound-effects side of this pack. It pulls in the MOSS-SoundEffect model (~3GB) from OpenMOSS and hands you a moss_se_model object you feed into MossSoundEffectGenerate. There's nothing clever about it - which is the point. The loaders in this pack do the fiddly work (downloads, device placement, precision) so the generate node can be a one-shot prompt box.

The inputs

• model_path - a single choice: OpenMOSS-Team/MOSS-SoundEffect. It auto-downloads to ComfyUI/models/moss_ttsd/ on first load if it isn't there, and the dropdown also lists anything you've dropped into that folder manually.
• device - auto / cuda / cpu / mps, default auto. auto picks CUDA if it's available, then MPS, then CPU, so most people never touch it.
• precision - fp32 (default) or bf16. Unlike the TTS loader, sound effects default to fp32 for stability; switch to bf16 to roughly halve the model's memory footprint if VRAM is tight. The generate node's code specifically suggests fp32 when you hit a CUDA device-side assert.
• moss_codec (optional) - wire the moss_codec output from MossAudioCodecLoadModel in here and the loader reuses the codec's local path instead of resolving the MOSS-Audio-Tokenizer itself. Do it; it saves a redundant download and keeps one codec copy shared across the graph.

How it loads

It uses HuggingFace's AutoModel + AutoProcessor with remote code enabled (MOSS models ship custom processor code from their repos), resolves the audio tokenizer path with a sensible priority order - connected codec node first, then a local copy in the models folder, then the default HuggingFace ID - and on a CUDA box with flash-attn installed it'll opt into flash_attention_2 for faster generation on Ampere-or-newer GPUs.

Install

ComfyUI Manager → search ComfyUI Kaola MOSS-TTS, or:

cd ComfyUI/custom_nodes
git clone https://github.com/kana112233/ComfyUI-kaola-moss-tts.git
cd ComfyUI-kaola-moss-tts
pip install -r requirements.txt

The one dependency that matters: transformers>=5.0.0. The pack's README calls it out as a hard requirement, and it can collide with packs pinned to older transformers - a fresh Python 3.12 environment is the recommended cure.

Where people get burned: mostly by expectations. This is a ~3GB model with all the attendant first-load download time, and loading it alongside other MOSS models in the same graph stacks VRAM quickly. If the graph is getting heavy, keep the codec shared and prefer bf16.