ComfyUI_RH_FlashTalk

ComfyUI custom nodes for SoulX-FlashTalk - Real-time infinite streaming of audio-driven avatars.

SoulX-FlashTalk is the first 14B model to achieve sub-second start-up latency (0.87s) while maintaining a real-time throughput of 32 FPS.

✨ Features

• Audio-Driven Generation: Generate talking face videos from audio input
• High Quality Output: Powered by SoulX-FlashTalk-14B model (14 billion parameters)
• Streaming Processing: Efficient chunk-based audio processing for long audio files
• Quantization Support: INT8 quantization for reduced VRAM usage

🛠️ Installation

1. Clone this repository into your ComfyUI custom_nodes folder:

cd ComfyUI/custom_nodes
git clone https://github.com/HM-RunningHub/ComfyUI_RH_FlashTalk.git

2. Install dependencies:

cd ComfyUI_RH_FlashTalk
pip install -r requirements.txt

3. Install Flash Attention (optional but recommended):

pip install ninja
pip install flash_attn==2.8.0.post2 --no-build-isolation

📦 Model Download

Download the following models and place them in your ComfyUI models folder:

SoulX-FlashTalk-14B

Download from HuggingFace: Soul-AILab/SoulX-FlashTalk-14B

huggingface-cli download Soul-AILab/SoulX-FlashTalk-14B --local-dir ComfyUI/models/Soul-AILab/SoulX-FlashTalk-14B

Place in: ComfyUI/models/Soul-AILab/SoulX-FlashTalk-14B/

Chinese Wav2Vec2

Download from HuggingFace: TencentGameMate/chinese-wav2vec2-base

huggingface-cli download TencentGameMate/chinese-wav2vec2-base --local-dir ComfyUI/models/wav2vec/chinese-wav2vec2-base

Place in: ComfyUI/models/wav2vec/chinese-wav2vec2-base/

🚀 Usage

Nodes

Basic Workflow

1. Add RunningHub FlashTalk Loader node to load the model
2. Connect the pipeline output to RunningHub FlashTalk Sampler
3. Provide an avatar image and audio input
4. Set seed for reproducibility
5. Queue the workflow to generate video

📝 Parameters

FlashTalk Sampler

Default Generation Settings

⚠️ Requirements

• CUDA-compatible GPU with sufficient VRAM (40GB+ recommended, use cpu_offload for lower VRAM)
• Python 3.10+
• FFmpeg (for video processing)

📚 Citation

If you find this work useful, please cite the original SoulX-FlashTalk paper:

@misc{shen2025soulxflashtalktechnicalreport,
      title={SoulX-FlashTalk: Real-Time Infinite Streaming of Audio-Driven Avatars via Self-Correcting Bidirectional Distillation}, 
      author={Le Shen and Qian Qiao and Tan Yu and Ke Zhou and Tianhang Yu and Yu Zhan and Zhenjie Wang and Ming Tao and Shunshun Yin and Siyuan Liu},
      year={2025},
      eprint={2512.23379},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.23379}, 
}

📄 License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

🙏 Acknowledgments

• Soul-AILab/SoulX-FlashTalk - The original FlashTalk model and inference code
• Alibaba Wan Team - Base model architecture
• TencentGameMate/chinese-wav2vec2-base - Audio feature extraction