Breeze TTS 2 · T8 模型加载器

Every Breeze TTS 2 workflow in this pack starts here, and most first-time failures happen here too. The T8_BreezeTTS_ModelLoader node loads the official BreezeBlue/Breeze-TTS-2 weights - sharded BF16 safetensors, pinned to a fixed revision - into ComfyUI's own memory management. The other five nodes are useless until this one produces a model, so treat it as the gate, not a formality.

The checkbox that bites everyone

accept_model_license defaults to false, and if you leave it there the node throws immediately. That's not a bug, it's the model's terms doing their job. Breeze TTS 2 ships under the BreezeBlue Research and Non-Commercial License - no commercial use, no non-consensual voice cloning. Read MODEL_LICENSE in the pack folder, understand what you're agreeing to, then flip the switch. If you're building anything you'd sell, stop right here and pick a permissively licensed TTS instead (the pack's own README is upfront about this).

What happens when you run it

With download_if_missing on (the default), the first run pulls the model from Hugging Face into ComfyUI/models/breeze_tts/BreezeBlue_Breeze-TTS-2. The revision is hard-pinned, so you won't wake up one day with different weights because the upstream repo moved - reproducibility is the whole point of a fixed revision. Downloads resume if interrupted, and the loader verifies the config, tokenizer, codec, the weights index, and every shard the index references before it will hand you a model. If a file comes back corrupted, the error names the exact file; delete just that one and re-run rather than nuking the folder.

Three choices matter on a normal machine:

• dtype (auto / bf16 / fp32) - auto picks bf16 on a capable GPU, which is what you want. Drop to fp32 if you're on a card that can't do bf16 or you're CPU-bound.
• device (auto / cuda / cpu) - auto means CUDA when it's there.
• attention (auto / eager / sdpa / flash_attention / sageattention) - auto is fine. sdpa is the safe fast default; flash_attention and sageattention need their own kernels installed, so don't pick them just because they sound fast.

Why this pack doesn't wreck your ComfyUI

The classic way an audio node pack ruins your weekend is by pinning its own torch/transformers and silently breaking the host environment - the KB's audio essay calls dependency conflict "the default failure mode" for TTS-in-ComfyUI. This pack goes the other way. Its requirements are deliberately light (accelerate, huggingface_hub, packaging, safetensors, soundfile, tqdm - no torch, no transformers, no numpy), and it runs a compatibility check on load. It needs transformers >= 4.57, < 6 (tested on 4.57.3 and 5.16.1); if yours is out of range it says so at registration instead of papering over it. So if you hit a "Transformers X is not supported" error, fix the host's transformers version - don't install a second copy for this node.

Install, and what you get out

Install via ComfyUI Manager (search "Breeze TTS 2 · T8star-Aix") or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-breeze-tts.git
python -m pip install -r Comfyui-breeze-tts/requirements.txt

You'll need ComfyUI 0.33.0+ and Python 3.10+. The outputs are model (type BREEZE_T8_MODEL, wired into the Generate node) and model_info, a JSON string that tells you exactly what loaded - model directory, revision, device, dtype, attention. Handy when you're chasing a weird result and want to confirm it's actually running bf16 on CUDA and not fp32 on CPU.

One more thing: the loader is not the fun part, but it's the part that downloads a real model on first run, so budget a few minutes and a decent connection. If it OOMs, the error suggests the right fix: stop other workflows or switch device to cpu. No torch reinstall required - this node won't touch it, and you shouldn't either.