MOSS-TTS Load Model

Every node in the MOSS-TTS 1.5 pack is useless until this one runs, because MOSS isn't a tiny ONNX model - it's a real transformer that wants to live in your VRAM. MOSSLoadModel is the gatekeeper: it pulls the weights from Hugging Face on first run, loads the processor + model into memory, and hands the whole bundle out as moss_model. Fan that output into the Speak, Voice Clone, and Voice Continue nodes and you're in business. The pack is a thin, MIT wrapper around OpenMOSS's Apache-2.0 model - no API, no key, no account. It's the rare local TTS that does multilingual cloning without any cloud call, which is why it also shows up bundled inside Blender filmmaking addons.

Two models, one dropdown

model_id defaults to MOSS-TTS-Local-Transformer-v1.5 (1.7B), a MossTTSLocal architecture that outputs 48 kHz stereo and needs ~12 GB VRAM in bfloat16. The other choice is the full MOSS-TTS-v1.5 (8B), a deeper MossTTSDelay model at 24 kHz that wants ~22 GB and can sound a bit more expressive. Same API, same 31 languages, same duration semantics, so you can swap without rewiring. On an RTX 3090 or 5090 both fit - but start with the 1.7B. It's faster, it leaves room for the video half of your pipeline, and the quality gap is subtle. The "(1.7B)" / "(8B)" suffix is just a UI label; it gets stripped before the Hugging Face call.

The other knobs

• device - cuda by default, falls back to cpu automatically if CUDA's missing. CPU works but is roughly 50x slower, so treat it as an emergency.
• attention - the one that used to bite people. MOSS's own model code defaults to flash_attention_2, which hard-crashes if the flash_attn package isn't installed. The auto default (recommended) checks whether flash-attn is present and falls back to PyTorch's built-in sdpa if not - so a plain install just works. Force flash_attention_2 only if you've actually installed flash-attn; on brand-new CUDA toolkits without a prebuilt Windows wheel, stay on sdpa. Quality is identical, only long-context speed differs.

dtype is picked for you: bfloat16 on CUDA (MOSS's training precision), float32 on CPU. The bundle caches by (model_id, device, attention), so re-queued workflows reuse the loaded model instead of paying the load cost every time. Output is a single moss_model that feeds any of the three generators.

Install and the first-run tax

Pack install is standard - Manager, search "MOSS-TTS 1.5", or:

cd ComfyUI/custom_nodes
git clone https://github.com/eehrich/ComfyUI-MOSS-TTS-1.5.git MOSS-TTS-ComfyUI

Restart ComfyUI. There are deliberately no pip dependencies - it rides on ComfyUI's own transformers (>= 4.50.3), and the current model build runs on both transformers 4.x and 5.x. The real cost is the first execution: ~9.1 GB for the 1.7B, ~17 GB for the 8B, streamed into your Hugging Face cache. Start the download and go make coffee.

The gotchas that actually get people

• Python 3.11+ dataclass crash. MOSS's audio-tokenizer config files declare fields without defaults after the parent added defaulted ones, so the first load dies with TypeError: non-default argument 'sampling_rate' follows default argument 'problem_type'. Fix once in the cached file under ~/.cache/huggingface/modules/transformers_modules/.../configuration_moss_audio_tokenizer.py by giving the fields = None defaults. Nothing behavioural changes - the real defaults still come from __init__.
• MODALITY_TO_BASE_CLASS_MAPPING AttributeError on transformers 4.x - that's a stale cached MOSS build from before the 4.x/5.x compatibility guard. Delete ~/.cache/huggingface/hub/models--OpenMOSS-Team--MOSS-TTS-* and let it re-download, or upgrade transformers to >= 5.0 (needs Python 3.10+).
• Can't load the model … pytorch_model.bin - the safetensors download stalled, usually low disk space in the HF cache. Rerun the download explicitly and free space.
• std::bad_alloc on import torchcodec - version mismatch between torchcodec and torch. MOSS itself doesn't need torchcodec, so uninstall it or match versions.

One VRAM note before you build a big scene: reference and prefix audio add roughly 1 GB per 20 s of history on top of the 12 GB baseline, so a very long reference clip can OOM you before generation even starts. Keep reference clips in the 5–15 s sweet spot and watch nvidia-smi during long continuation chains.