title: catvton-flux emoji: 🖥️ colorFrom: yellow colorTo: pink sdk: gradio sdk_version: 5.0.1 app_file: app.py pinned: false license: mit

catvton-flux

An state-of-the-art virtual try-on solution that combines the power of CATVTON (CatVTON: Concatenation Is All You Need for Virtual Try-On with Diffusion Models) with Flux fill inpainting model for realistic and accurate clothing transfer. Also inspired by In-Context LoRA for prompt engineering.

Update

Latest Achievement

(2024/11/26):

Updated the weights. (Still training on the VITON-HD dataset only.)
- Reduce the fine-tuning weights size (46GB -> 23GB)
- Weights has better performance on garment small details/text.

(2024/11/25):

Released lora weights. Lora weights achieved FID: 6.0675811767578125 on VITON-HD dataset. Test configuration: scale 30, step 30.
Revise gradio demo. Added huggingface spaces support.
Clean up the requirements.txt.

(2024/11/24):

Released FID score and gradio demo
CatVton-Flux-Alpha achieved SOTA performance with FID: 5.593255043029785 on VITON-HD dataset. Test configuration: scale 30, step 30. My VITON-HD test inferencing results available here

Showcase

Original	Garment	Result

Model Weights

Fine-tuning weights in Hugging Face: 🤗 catvton-flux-alpha

LORA weights in Hugging Face: 🤗 catvton-flux-lora-alpha

The model weights are trained on the VITON-HD dataset.

Prerequisites

Make sure you are running the code with VRAM >= 40GB. (I run all my experiments on a 80GB GPU, lower VRAM will cause OOM error. Will support lower VRAM in the future.)

bash
conda create -n flux python=3.10
conda activate flux
pip install -r requirements.txt
huggingface-cli login

Usage

Run the following command to try on an image:

LORA version:

python tryon_inference_lora.py \
--image ./example/person/00008_00.jpg \
--mask ./example/person/00008_00_mask.png \
--garment ./example/garment/00034_00.jpg \
--seed 4096 \
--output_tryon test_lora.png \
--steps 30

Fine-tuning version:

python tryon_inference.py \
--image ./example/person/00008_00.jpg \
--mask ./example/person/00008_00_mask.png \
--garment ./example/garment/00034_00.jpg \
--seed 42 \
--output_tryon test.png \
--steps 30

Run the following command to start a gradio demo with LoRA weights:

python app.py

Run the following command to start a gradio demo without LoRA weights:

python app_no_lora.py

Gradio demo:

TODO:

Release the FID score
Add gradio demo
Release updated weights with better performance
Train a smaller model
Support comfyui

Citation

@misc{chong2024catvtonconcatenationneedvirtual,
 title={CatVTON: Concatenation Is All You Need for Virtual Try-On with Diffusion Models}, 
 author={Zheng Chong and Xiao Dong and Haoxiang Li and Shiyue Zhang and Wenqing Zhang and Xujie Zhang and Hanqing Zhao and Xiaodan Liang},
 year={2024},
 eprint={2407.15886},
 archivePrefix={arXiv},
 primaryClass={cs.CV},
 url={https://arxiv.org/abs/2407.15886}, 
}
@article{lhhuang2024iclora,
  title={In-Context LoRA for Diffusion Transformers},
  author={Huang, Lianghua and Wang, Wei and Wu, Zhi-Fan and Shi, Yupeng and Dou, Huanzhang and Liang, Chen and Feng, Yutong and Liu, Yu and Zhou, Jingren},
  journal={arXiv preprint arxiv:2410.23775},
  year={2024}
}

Thanks to Jim for insisting on spatial concatenation. Thanks to dingkang MoonBlvd Stevada for the helpful discussions.

License

The code is licensed under the MIT License.
The model weights have the same license as Flux.1 Fill and VITON-HD.