MOSS-TTS v1.5 Load Model

Every MOSS-TTS workflow starts here. This is the node that picks which model loads, how heavy it runs, and where its 9+ GB of weights come from - and it's the node that proves the pack author actually thought about the ComfyUI ecosystem instead of just wrapping a Gradio demo.

The pack wraps two real variants of OpenMOSS's MOSS-TTS v1.5, both by OpenMOSS-Team under Apache-2.0, and you choose between them with the model dropdown:

• Local-Transformer - a Qwen3-4B-class backbone with a nano-GPT2 local transformer, 48 kHz stereo output, ~9.1 GB of bf16 weights. This is the default and the one most people will actually use.
• Delay - the 8B delay-pattern model, 24 kHz, roughly double the VRAM.

One quick correction while you're here: some READMEs floating around call the local backbone "1.7B." That's wrong - the config is Qwen3-4B-shaped and the checkpoint runs about 9.1 GB in bf16. Trust the numbers you see here, not the copy-paste.

What you actually set

Four inputs, and only two of them matter on day one.

• model - pick your variant. Defaults to Local-Transformer, which is the right call unless you specifically need the 8B's long-form behavior.
• dtype - auto means bf16 on CUDA, fp32 on CPU. Leave it on auto unless you have a reason not to.
• attention - auto picks flash_attention_2 if the flash_attn package is installed, otherwise sdpa. And here's the thing: sdpa quality is identical. If you ever get flash-attn errors, this is the toggle that fixes it.
• download_if_missing - leave it on. It only downloads when the weights aren't already local.

The single output is mosstts_model, which plugs into Generate Speech, Voice Clone, or Continue Speech.

How loading actually works (this is the good part)

Two things make this pack nicer than most TTS-in-ComfyUI. First, no trust_remote_code: the model's code is vendored inside the pack under assets/ with small, clearly-marked compatibility patches, so it doesn't depend on whatever random version of the model code happens to be sitting in your HF module cache. Second, it runs on both transformers 4.x and 5.x, which is rare in the TTS corner - most audio node packs fight transformers conflicts as a hobby (the KB's audio essay calls dependency conflict "the default failure mode" of TTS in ComfyUI; this pack dodges that whole category).

Weights resolve in order: $MOSS_TTS_MODELS_DIR/<Repo-Name> → ComfyUI/models/mosstts/<Repo-Name> → your HF hub cache → download. Downloads are chunky - about 9.1 GB + ~8 GB codec for Local, ~17 GB + ~6.7 GB for Delay - so the first load is a "go make tea" moment. VRAM-wise, plan on ~12 GB bf16 for Local and ~22 GB for Delay.

The weights also register through ComfyUI's memory management (ModelPatcher/ModelPatcherDynamic), so unloading and Free memory work like they do for checkpoints. When your video workflow and this TTS fight for the same card, you can actually free the model.

Install

ComfyUI Manager → Custom Nodes Manager → search moss-tts → Install, then restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/CloudRipple/MOSS-TTS-ComfyUI.git
python install.py   # only installs what's missing

That's the other nice surprise: the dependency list is just huggingface-hub, safetensors, numpy, tqdm. Torch, torchaudio, and transformers come from your ComfyUI install - this pack deliberately doesn't pin its own, which is exactly why it doesn't blow up your other nodes.

When it goes wrong

The two failure modes worth knowing:

• "Model can't be found" / nothing happens - the loader prints the exact search paths it tried. Point $MOSS_TTS_MODELS_DIR at a directory containing the repo folders, or make sure HF_HOME points at your cache.
• flash-attn errors on load - switch attention to sdpa or auto. Flash attention is optional here, not required.

The only real gotcha is patience: that first download is large and there's no progress bar in the node title itself. Let it run.