ComfyUI-Qwen2_5-Omni

Qwen2.5-Omni in ComfyUI

License: Apache License 2.0 from Qwen2.5-Omni

Qwen2.5-Omni, the new flagship end-to-end multimodal model in the Qwen series. Designed for comprehensive multimodal perception, it seamlessly processes diverse inputs including text, images, audio, and video, while delivering real-time streaming responses through both text generation and natural speech synthesis.

WIP: 4-bit quantized Qwen2.5-Omni-7B (GPTQ-Int4/AWQ) support

Supports Type:

• Image
• Multi image
• Video
• Audio
• Text only
• Mixed type

Please refer to the following example image:

Qwen2.5-Omni

Qwen2.5-Omni

Qwen2.5-Omni

Qwen2.5-Omni

Qwen2.5-Omni

Installation:

Clone this repository to 'ComfyUI/custom_nodes` folder.

Install the dependencies in requirements.txt, transformers version 4.52.3 minimum is required:

pip install -r requirements.txt

or if you use portable (run this in ComfyUI_windows_portable -folder):

python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-Qwen2_5-Omni\requirements.txt

Supports most Qwen2_5-Omni models, which can be automatically downloaded with the DownloadAndLoadQWEN2_5_OMNIModel to ComfyUI/models/VLM:

"Qwen/Qwen2.5-Omni-7B",
"Qwen/Qwen2.5-Omni-3B",

node name：

• Qwen2_5-Omni/DownloadAndLoadQWEN2_5_OMNIModel
• Qwen2_5-Omni/QWEN2_5_OMNI_Run
• Qwen2_5-Omni/Qwen2_5_OmniBatchImage

Citation

@article{Qwen2.5-Omni,
  title={Qwen2.5-Omni Technical Report},
  author={Jin Xu, Zhifang Guo, Jinzheng He, Hangrui Hu, Ting He, Shuai Bai, Keqin Chen, Jialin Wang, Yang Fan, Kai Dang, Bin Zhang, Xiong Wang, Yunfei Chu, Junyang Lin},
  journal={arXiv preprint arXiv:2503.20215},
  year={2025}
}