This project provides a ComfyUI wrapper of FLOAT for Generative Motion Latent Flow Matching for Audio-driven Talking Portrait

For a more advanced and maintained version, check out: ComfyUI-FLOAT_Optimized

Last tested: 2 January 2026 (ComfyUI v0.7.0@f2fda02 | Torch 2.9.1 | Python 3.10.12 | RTX4090 | CUDA 13.0 | Debian 12)

⭐ Support

If you like my projects and wish to see updates and new features, please consider supporting me. It helps a lot!

[](https://github.com/yuvraj108c/ComfyUI-Depth-Anything-Tensorrt)

[](https://github.com/yuvraj108c/ComfyUI-Upscaler-Tensorrt)

[](https://github.com/yuvraj108c/ComfyUI-Dwpose-Tensorrt)

[](https://github.com/yuvraj108c/ComfyUI-Rife-Tensorrt)

[](https://github.com/yuvraj108c/ComfyUI-Whisper)

[](https://github.com/yuvraj108c/ComfyUI_InvSR)

[](https://github.com/yuvraj108c/ComfyUI-FLOAT)

[](https://github.com/yuvraj108c/ComfyUI-Thera)

[](https://github.com/yuvraj108c/ComfyUI-Video-Depth-Anything)

[](https://github.com/yuvraj108c/ComfyUI-PiperTTS)

[](https://www.buymeacoffee.com/yuvraj108cZ)

[](https://paypal.me/yuvraj108c)

---

🚀 Installation

git clone https://github.com/yuvraj108c/ComfyUI-FLOAT.git
cd ./ComfyUI-FLOAT
pip install -r requirements.txt

☀️ Usage

Load example workflow

Upload driving image and audio, click queue

Models autodownload to /ComfyUI/models/float

The models are organized as follows:

“`.bash

|– float.pth # main model

|– wav2vec2-base-960h/ # audio encoder

| |– config.json

| |– model.safetensors

| |– preprocessor_config.json

|– wav2vec-english-speech-emotion-recognition/ # emotion encoder

|– config.json

|– preprocessor_config.json

|– pytorch_model.bin

🛠️ Parameters

ref_image: Reference image with a face (must have batch size 1)

ref_audio: Reference audio (For long audios (e.g 3+ minutes), ensure that you have enough ram/vram)

a_cfg_scale: Audio classifier-free guidance scale (default:2)

r_cfg_scale: Reference classifier-free guidance scale (default:1)

emotion: none, angry, disgust, fear, happy, neutral, sad, surprise (default:none)

e_cfg_scale: Intensity of emotion (default:1). For more emotion intensive video, try large value from 5 to 10

crop: Enable only if the reference image does not have a centered face

fps: Frame rate of the output video (default:25)

Citation

@article{ki2024float,
  title={FLOAT: Generative Motion Latent Flow Matching for Audio-driven Talking Portrait},
  author={Ki, Taekyung and Min, Dongchan and Chae, Gyeongsu},
  journal={arXiv preprint arXiv:2412.01064},
  year={2024}
}

Acknowledgments

Thanks to simplepod.ai for providing GPU servers

License

Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)