This is the official implementation of the paper:

OmniVTON: Training-Free Universal Virtual Try-On
Zhaotong Yang, Yuhui Li, Shengfeng He, Xinzhe Li, Yangyang Xu, Junyu Dong, Yong Du^* （ * indicates corresponding author)
Proceedings of the International Conference on Computer Vision

If interested, star 🌟 this project!

To set up the environment, install requirements using Python 3.9:

conda create -n omnivton python=3.9
conda activate omnivton
pip install -r requirements.txt

Generate agnostic mask images following the operation provided by CAT-DM.

Use the clip-interrogator to obtain text prompts for both the target garment image and the original garment image.

• Note: If the original garment image is unavailable, you may extract it using a semantic segmentation model and then feed the result into the clip-interrogator.

Structure of the condition folder should be as follows:

condition dir
|-- image_clip_interrogate       # original garment prompt
    |-- xxx_ci.json
|-- cloth_clip_interrogate       # target garment prompt
    |-- xxx_ci.json

To generate a pseudo person image, run the following command:

python3.9 inference.py \
  --model-id sd2_inp \
  --image-path PERSON_IMAGE_PATH \
  --cloth-path CLOTHING_IMAGE_PATH \
  --in-mask-path AGNOSTIC_MASK_PATH \
  --out-mask-path CLOTHING_MASK_PATH \
  --condition-path CONDITION_DIR_PATH \
  --stage outpainting --H 512 --W 384 \
  --output-path OUTPUT_PATH

--model-id specifies the baseline model for virtual try-on. The following baseline models are supported by the script:

• sd2_inp - Stable Diffusion 2.0 Inpainting
• sd15_inp - Stable Diffusion 1.5 Inpainting

If not specified --model-id defaults to sd2_inp.

The script uses a combination of positive and negative prompts by default to achieve more visually appealing results.

On the first run, it will automatically download the required models — please be patient. Alternatively, you may choose to load models from a local directory.

Once the pseudo person image is obtained, we need to extract both human part segmentation map and 25 keypoint coordinates for the pseudo person image and the original person image, using TAPPS and OpenPose, respectively.

• Note: To facilitate the extraction of semantic maps, please replace the files in the official TAPPS directory with those provided in the tapps folder of this repository.

Then, the condition folder should be updated as follows:

condition dir
|-- image_clip_interrogate       # original garment prompt
    |-- xxx_ci.json
|-- image_tapps_parse            # original person segmentation map
    |-- xxx_pps.png
|-- image_openpose_json          # original person keypoints
    |-- xxx_keypoints.json
|-- cloth_clip_interrogate       # target garment prompt
    |-- xxx_ci.json
|-- cloth_tapps_parse            # pseudo person segmentation map
    |-- xxx_pps.png
|-- cloth_openpose_json          # pseudo person keypoints
    |-- xxx_keypoints.json

To generate the final try-on result, run the following command:

python3.9 inference.py \
  --model-id sd2_inp \
  --image-path PERSON_IMAGE_PATH \
  --cloth-path CLOTHING_IMAGE_PATH \
  --in-mask-path PERSON_MASK_PATH \
  --out-mask-path CLOTHING_MASK_PATH \
  --condition-path CONDITION_DIR_PATH \
  --stage vton --H 512 --W 384 \
  --output-path OUTPUT_PATH

Thanks to HD-Painter for the base code.

Thanks to CAT-DM for the masking generation code.

Thanks to TAPPS and OpenPose for human parsing.

If you find our work useful for your research, please cite us:

@article{yang2025omnivton,
  title={OmniVTON: Training-Free Universal Virtual Try-On},
  author={Yang, Zhaotong and Li, Yuhui and He, Shengfeng and Li, Xinzhe and Xu, Yangyang and Dong, Junyu and Du, Yong},
  journal={arXiv preprint arXiv:2507.15037},
  year={2025}
}