MOSS-TTS Load Model

MOSS-TTS Load Model is the front door to the whole pack. It loads OpenMOSS's MOSS-TTS Local Transformer v1.5 plus its stereo audio tokenizer into a single mosstts_model bundle, and every other node in this pack - Generate Speech, Voice Clone, Continue Speech - takes that bundle as its first input. No loader, nothing else lights up. It's also the only node here you actually configure, because it's the one that decides how much VRAM the TTS stack eats and where the weights come from.

Why you'd bother: MOSS-TTS v1.5 is a local, 48 kHz stereo Transformer TTS that does direct generation, zero-shot voice cloning, and continuation in 31 languages, and in the 2026 local-TTS scene it's the model people single out for expression - more natural delivery than a lot of the "just clone the timbre" crowd. All of that lives behind this one loader.

How it works

The pack bundles the small remote-code, config, and tokenizer files in its own assets/ folder. The loader links those beside the actual weights in ComfyUI/models/mosstts/ so Transformers can load from a normal local folder with trust_remote_code=True. What it actually downloads - when download_if_missing is on - is the main model.safetensors plus the three-shard MOSS-Audio-Tokenizer-v2, both from Hugging Face.

The bundle is cached. Load once, run dozens of generations, and you keep the model resident; change the model, dtype, attention backend, or tokenizer files and it unloads the old bundle internally. There's no separate unload node, which is a small relief. If you run ComfyUI with AIMDO DynamicVRAM, the loader registers the model, codec, and optional Whisper ASR with native dynamic patchers so memory gets tracked properly.

The inputs that matter

Only four, and you'll realistically touch two:

• model - one catalog entry for v1.5 (BF16, auto-download). Nothing to choose from yet.
• dtype - auto reads the config (currently BF16). Manual bf16 / fp16 are there if you're tuning memory.
• attention - auto prefers FlashAttention 2 when it's installed and compatible, falls back to SDPA on CUDA and eager on CPU. Leave it on auto. The README says the same: only touch this while debugging.
• download_if_missing - true by default. Flip it off for a fully offline setup after your first run.

Output: mosstts_model, which wires straight into the Voice Clone, Continue Speech, and Generate Speech nodes' mosstts_model input.

Installing

Same for every node in this pack. Easiest: ComfyUI Manager, search MOSS-TTS and install, then restart. Or via git URL:

https://github.com/Saganaki22/Moss_TTS-ComfyUI

Manual install from ComfyUI/custom_nodes:

git clone https://github.com/Saganaki22/Moss_TTS-ComfyUI.git
../venv/bin/python Moss_TTS-ComfyUI/install.py

install.py only installs lightweight missing deps - accelerate, huggingface-hub, numpy, safetensors, tqdm. It deliberately does not touch your torch, torchaudio, or transformers, so it won't clobber another node's stack (a real hazard in the ComfyUI ecosystem, where every node shares one Python environment). It does warn if your Transformers is below 4.57; v5 is what it recommends.

Common issues

The first load downloads a multi-gigabyte model plus the codec, so it looks frozen for a while. That's normal.

If the download lands on a mirror you don't want, it's another custom node mutating HF_ENDPOINT globally - this pack doesn't set it. The README documents two MOSS-specific errors - Unexpected keyword argument local_files_only and property 'device' ... has no setter - and the pack patches both internally; if you hit them, you're on an old version, update. And remember: with download_if_missing off, every file has to already sit at its path under ComfyUI/models/mosstts/ or the loader raises a clear FileNotFoundError telling you exactly which one's missing.