MagicTryOn: The New Standard in Video Virtual Try-On

A state-of-the-art framework that delivers unparalleled realism and stability, preserving garment details with remarkable precision. See how clothes truly look and move in dynamic video.

A collaborative research project by Zhejiang University & vivo

Get Started View on GitHub Read The Paper

Why MagicTryOn is a Game-Changer

Flawless Detail Preservation

From intricate patterns to delicate textures like lace, MagicTryOn perfectly retains the original appearance of the garment, ensuring a photorealistic result.

Unmatched Temporal Consistency

Even with significant body movements like dancing or turning, the virtual garment remains stable and moves naturally with the person, free from flicker or distortion.

Universal Applicability

Our framework is designed for maximum flexibility. It works with any person, any garment, and even in unconventional scenarios like virtual try-on for dolls, without specialized training.

Proven Performance Gains

30%

Higher Detail Accuracy

Significantly outperforms previous methods in preserving the fine details and textures of clothing items.

25%

Improved Temporal Consistency

Delivers a more stable and natural look across video frames, eliminating common visual artifacts.

2x

Faster Processing Speed

Achieves superior quality results at double the speed of comparable virtual try-on systems.

The Core Technology

Diffusion Transformer Backbone

We replace the standard U-Net with a powerful Diffusion Transformer (DiT) to better model spatiotemporal relationships, ensuring high-definition and consistent video output.

Coarse-to-Fine Preservation

A dual strategy that uses garment tokens for initial guidance and multi-level conditions (semantics, textures, lines) for fine-grained refinement during the denoising process.

Mask-Aware Loss

An innovative loss function that focuses the model's attention on the garment region, preventing background interference and maximizing the fidelity of the virtual try-on.

Get Started in 3 Steps

1. Set Up Environment

Use Conda to create an isolated environment with the required dependencies.

conda create -n magictryon python==3.12.9 conda activate magictryon pip install -r requirements.txt

2. Download Model Weights

Get the pretrained model from HuggingFace.

cd Magic-TryOn HF_ENDPOINT=https://hf-mirror.com huggingface-cli download LuckyLiGY/MagicTryOn --local-dir ./weights/MagicTryOn_14B_V1

3. Run the Demo

Execute the inference script for either image or video try-on.

# Example: Video try-on for upper body garments CUDA_VISIBLE_DEVICES=0 python predict_video_tryon_up.py

For more detailed instructions, including custom try-on, please visit our GitHub repository.

Citation

If you use MagicTryOn in your research, please cite our paper.

@misc{li2025magictryon,
      title={MagicTryOn: Harnessing Diffusion Transformer for Garment-Preserving Video Virtual Try-on}, 
      author={Guangyuan Li and Siming Zheng and Hao Zhang and Jinwei Chen and Junsheng Luan and Binkai Ou and Lei Zhao and Bo Li and Peng-Tao Jiang},
      year={2025},
      eprint={2505.21325},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2505.21325}, 
}