Load MOSS Voice Generator Model

MossVoiceGeneratorLoadModel loads the MOSS-VoiceGenerator model (~3GB) - the piece of this pack that creates a voice from a text description instead of a reference clip. The loader itself is unremarkable: pick the model, pick where it runs, pick the precision, and out pops a moss_vg_model object for MossVoiceGeneratorGenerate to use. But it has one genuinely useful habit worth understanding before you wire anything.

The inputs

• model_path - a single choice, OpenMOSS-Team/MOSS-VoiceGenerator, auto-downloaded to ComfyUI/models/moss_ttsd/ on first load. Anything you drop into that folder also appears in the dropdown.
• device - auto / cuda / cpu / mps, default auto. Resolves CUDA → MPS → CPU, so leave it alone unless you have a reason.
• precision - fp32 (default) or bf16. Defaults to fp32 for stability, matching the sound-effect loader; bf16 roughly halves the memory footprint if you need it. If generation throws a CUDA device-side assert, the generate node's own error message points you back here to force fp32.
• moss_codec (optional) - this is the input the node's description is begging you to use: connect the moss_codec output from MossAudioCodecLoadModel "to avoid redundant downloads/loading." The loader resolves the shared MOSS-Audio-Tokenizer in a priority order - your connected codec first, then a local copy, then the default HuggingFace ID - so feeding it the codec node means one tokenizer for the whole graph instead of one per model.

How it loads

Standard pattern for this pack: HuggingFace AutoModel + AutoProcessor with remote code, normalize_inputs=True on the processor, flash-attention probe on CUDA (flash_attention_2 on Ampere+, sdpa otherwise), model placed on the chosen device in eval mode. One detail that differs from the codec node: the audio tokenizer gets moved onto the GPU here, because the voice generator is the one actually doing the encoding work.

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

Standing requirements for the pack: transformers>=5.0.0 (a fresh Python 3.12 environment is the README's recommended fix for conflicts), and the ~3GB model download plus the ~1GB codec on first run.

Where people get burned: not connecting the codec and watching the loader pull the tokenizer separately, then again when the TTSD side loads it - three redundant downloads, all because one wire was skipped. Load the codec once, fan it out, and save yourself the bandwidth.