Qwen3-TTS Loader

Qwen3-TTS is Alibaba's flagship speech model, and the fun part is that it isn't one model - it's three flavors that do three different jobs. This loader is how you pick which one lands in your graph. One of them clones a voice from three seconds of audio. One ships with Alibaba's nine built-in voices. One lets you describe a voice into existence ("raspy deep male, slightly breathy") and it builds it. Same pipe-style output, wildly different behaviors, and this node is the dropdown that decides which you get.

The three model families

The model_name dropdown (five options) breaks down like this:

• Base (1.7B or 0.6B) - the cloning workhorse. 3-second zero-shot voice clone, and per the README's own table it's the strongest clone of the family. No emotion control of its own beyond what's in the recording.
• CustomVoice - Alibaba's built-in presets (Vivian, Serena, Uncle_Fu, and the rest) plus emotion and dialect control. The clone quality is explicitly "very poor" here - don't bring your reference audio to a preset party.
• VoiceDesign (1.7B) - the "make up a voice from a description" engine. No cloning, but the best emotion and dialect handling of the three.

The inputs

• model_name - the enum above. Default is Qwen/Qwen3-TTS-12Hz-1.7B-Base.
• device - auto is the default; cuda, cpu, and mps are there if you want to force it.
• dtype - bf16 by default, and on anything newer than a 10-series card that's the right choice.
• local_path (optional) - if you pre-downloaded the model (or HF is blocked where you are), point it at the folder instead of letting the node fetch it.

Output: qwen_model (QWEN_MODEL), which feeds the Qwen3-TTS Synthesis node.

Install

The pack install is the standard two-step:

cd ComfyUI/custom_nodes
git clone https://github.com/havvk/ComfyUI_AIIA.git
# restart ComfyUI

The first time you run the loader it auto-installs the qwen-tts Python package (and patches it for compatibility with current transformers - the official package breaks on newer releases, so this is genuinely useful, not theater). Weights auto-download from HuggingFace into ComfyUI/models/qwen_tts/. If that stalls, fetch manually:

cd ComfyUI/models
hf download Qwen/Qwen3-TTS-12Hz-1.7B-Base --local-dir qwen_tts/Qwen/Qwen3-TTS-12Hz-1.7B-Base

Gotchas

• The auto-install is a "surprise pip in your ComfyUI" moment the first time - fine, but it means your first load can take a couple of minutes while packages land.
• Flash Attention 2 is "strongly recommended" by the README for acceptable speed on the 1.7B models; without it, expect patience.
• If you connect a local_path, the node uses it directly and skips the auto-download - handy offline, but you own the folder layout then.