Extensions/ComfyUI-Qwen3-TTS
ComfyUI Extension

ComfyUI-Qwen3-TTS

ComfyUI-Qwen3-TTS brings Alibaba's powerful Qwen3-TTS models to ComfyUI Multi-GPU Support: CUDA, Apple Silicon (MPS), Intel Arc (XPU), and CPU

By ai-joe-git·Created 7 months ago·Updated 6 months ago· 5
ai-joe-git/ComfyUI-Qwen3-TTS
Nodes—
On cloudLocal install
Stars5
Updated6 months ago
Readme

ComfyUI-Qwen3-TTS

<div align="center">

Version License ComfyUI Python

High-quality Text-to-Speech nodes for ComfyUI using Qwen3-TTS models

Features • Installation • Usage • GPU Support • Examples

</div>

📖 Overview

ComfyUI-Qwen3-TTS brings Alibaba's powerful Qwen3-TTS models to ComfyUI, enabling high-quality text-to-speech synthesis with three distinct capabilities:

  • 🎭 Voice Cloning: Clone any voice from a reference audio sample
  • 🎨 Voice Design: Create custom voices from text descriptions
  • 🎵 Custom Voice: Use preset high-quality voices

Based on the open-source Qwen3-TTS project by Alibaba Qwen team.

✨ Features

Three Powerful Nodes

🎭 Qwen3-TTS VoiceClone

Clone voices from reference audio and synthesize new speech:

  • Upload any audio file as voice reference
  • Optional: Provide reference text for better accuracy
  • Generate speech in the cloned voice
  • Supports 11 languages
  • X-vector only mode for quick voice embedding extraction

🎨 Qwen3-TTS VoiceDesign

Create custom voices from natural language descriptions:

  • Describe the voice you want (e.g., "A deep male voice with a slight accent")
  • Model generates speech matching your description
  • Highly experimental and creative
  • Only available with 1.7B model

🎵 Qwen3-TTS CustomVoice

Use preset high-quality voices:

  • 9 preset speakers: Aiden, Dylan, Eric, Ono_anna, Ryan, Serena, Sohee, Uncle_fu, Vivian
  • Optional style instructions for voice modulation
  • Consistent, reliable quality
  • Available in both 0.6B and 1.7B models

Key Capabilities

✅ Multi-GPU Support: CUDA, Apple Silicon (MPS), Intel Arc (XPU), and CPU
✅ Automatic Model Download: Models are downloaded automatically on first run
✅ Precision Options: BFloat16 and Float32 for speed/quality trade-offs
✅ 11 Languages: Auto-detect or manually select from Chinese, English, Japanese, Korean, French, German, Spanish, Portuguese, Russian, Italian
✅ Flexible Model Sizes: 0.6B for speed, 1.7B for quality
✅ Deterministic Generation: Seed control for reproducible results

📦 Installation

Method 1: Manual Installation

cd ComfyUI/custom_nodes/
git clone https://github.com/ai-joe-git/ComfyUI-Qwen3-TTS.git
cd ComfyUI-Qwen3-TTS
pip install -r requirements.txt

Method 2: Git Submodule

If your ComfyUI is a git repository:

cd ComfyUI/custom_nodes/
git submodule add https://github.com/ai-joe-git/ComfyUI-Qwen3-TTS.git
cd ComfyUI-Qwen3-TTS
pip install -r requirements.txt

Important Notes

⚠️ First Run: Models (~6GB) will be automatically downloaded to ComfyUI/models/qwen-tts/
⚠️ Dependencies: Requires PyTorch 2.0+ and transformers 4.30+
⚠️ Qwen TTS Package: The qwen_tts package should be included as a submodule or folder

🎮 Usage

Basic Workflow

  1. Add a Qwen3-TTS node from the Qwen3-TTS category
  2. Configure settings:
    • Device: auto (recommended), cuda, mps, xpu, or cpu
    • Precision: bf16 (faster) or fp32 (more compatible)
    • Model: 0.6B (faster) or 1.7B (better quality)
  3. Enter your text
  4. Run the workflow

Node-Specific Usage

Voice Clone Node

Inputs:
- ref_audio: Reference audio (use Load Audio node)
- ref_text: Optional transcription of reference audio
- target_text: Text to synthesize in cloned voice
- device: auto/cuda/mps/xpu/cpu
- precision: bf16/fp32
- model_choice: 0.6B/1.7B
- language: Auto or specific language
- seed: For reproducible results
- x_vector_only: Extract voice embedding only (faster)

Voice Design Node

Inputs:
- text: Text to synthesize
- instruct: Voice description (e.g., "A cheerful female voice")
- device: auto/cuda/mps/xpu/cpu
- precision: bf16/fp32
- model_choice: 1.7B only
- language: Auto or specific language
- seed: For reproducible results

Custom Voice Node

Inputs:
- text: Text to synthesize
- speaker: Choose from 9 preset voices
- device: auto/cuda/mps/xpu/cpu
- precision: bf16/fp32
- model_choice: 0.6B/1.7B
- language: Auto or specific language
- instruct: Optional style instruction
- seed: For reproducible results

🖥️ GPU Support

CUDA (NVIDIA GPUs)

✅ Recommended: Best performance
✅ Auto-detected: Set device to auto
✅ Precision: BFloat16 recommended

Apple Silicon (MPS)

✅ Supported: M1/M2/M3 Macs
✅ Auto-detected: Set device to auto
⚠️ Note: Automatically uses FP16/BF16 for better performance

Intel Arc (XPU)

✅ Supported: Arc A-series discrete GPUs and iGPUs
⚠️ Performance Note: May be slower than CPU for small models
💡 Optimization: Install Intel Extension for PyTorch:

pip install intel-extension-for-pytorch

Expected Performance (Intel Arc):

  • CPU: Often faster for TTS workloads
  • XPU without IPEX: 2-5x slower than CPU
  • XPU with IPEX: Competitive with CPU

Recommendation: Use device="cpu" for Intel Arc iGPUs unless you have IPEX installed

CPU

✅ Universal fallback: Works on all systems
✅ Good performance: Especially for smaller models
✅ Recommended for: Intel Arc iGPUs (without IPEX)

📊 Examples

Example 1: Voice Cloning

1. Load reference audio → Qwen3-TTS VoiceClone
2. Set target_text: "Hello, this is a test of voice cloning"
3. Set device: "auto", precision: "bf16", model: "1.7B"
4. Output → Save Audio

Example 2: Custom Voice with Specific Style

1. Qwen3-TTS CustomVoice
2. Set speaker: "Ryan"
3. Set text: "Welcome to the future of AI"
4. Set instruct: "Speak with excitement and energy"
5. Set device: "auto", precision: "bf16"
6. Output → Save Audio

Example 3: Voice Design

1. Qwen3-TTS VoiceDesign
2. Set text: "Good morning everyone"
3. Set instruct: "A warm, elderly grandfather voice with a slight rasp"
4. Set device: "cuda", precision: "bf16", model: "1.7B"
5. Output → Save Audio

🔧 Troubleshooting

Models Not Downloading

# Manually download models
from huggingface_hub import snapshot_download
snapshot_download(repo_id="Qwen/Qwen3-TTS-12Hz-1.7B-Base", 
                  local_dir="ComfyUI/models/qwen-tts/Qwen3-TTS-12Hz-1.7B-Base")

Import Errors

# Reinstall dependencies
cd ComfyUI/custom_nodes/ComfyUI-Qwen3-TTS
pip install -r requirements.txt --force-reinstall

Slow Performance on Intel Arc GPU

# Install IPEX for better performance
pip install intel-extension-for-pytorch

# Or use CPU instead
# Set device="cpu" in the node

Out of Memory

  • Use 0.6B model instead of 1.7B
  • Set device to CPU
  • Use BF16 precision
  • Reduce max_new_tokens

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

  1. Fork the repository
  2. Create your feature branch (git checkout -b feature/AmazingFeature)
  3. Commit your changes (git commit -m 'Add some AmazingFeature')
  4. Push to the branch (git push origin feature/AmazingFeature)
  5. Open a Pull Request

📝 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

  • Alibaba Qwen Team for the Qwen3-TTS models
  • ComfyUI for the amazing UI framework
  • All contributors and users of this project

📮 Support

🗺️ Roadmap

  • [ ] Add streaming audio generation
  • [ ] Support for custom fine-tuned models
  • [ ] Batch processing support
  • [ ] Audio preview in ComfyUI
  • [ ] More preset voices
<div align="center">

Star ⭐ this repo if you find it useful!

Made with ❤️ for the ComfyUI community

</div>