---
license: cc-by-nc-4.0
library_name: decovton
tags:
  - virtual-try-on
  - diffusion
  - stable-diffusion
  - arxiv:2511.18775
datasets:
  - VITON-HD
  - DressCode
base_model: stable-diffusion-v1-5/stable-diffusion-inpainting
---

# DeCo-VTON

UNet checkpoints for **DeCo-VTON**, the official implementation of
**Rethinking Garment Conditioning in Diffusion-based Virtual Try-On: Decouple,
Don't Denoise**, accepted at ECCV 2026.

DeCo-VTON is a single-UNet virtual try-on model that separates garment
conditioning from the denoising target.

- Code: https://github.com/Levinna/DeCo-VTON
- Paper: https://arxiv.org/abs/2511.18775

This repository was formerly named `levinna/Re-CatVTON`. The old repository URL
redirects here, and the original checkpoint paths remain available for backward
compatibility.

## Checkpoints

| Subfolder | Dataset | Resolution | Availability |
|---|---|---|---|
| `VITON-HD-512/unet` | VITON-HD | 512×384 | Available |
| `DressCode-512/unet` | DressCode | 512×384 | Available |
| `VITON-HD-1024/unet` | VITON-HD | 1024×768 | Checkpoint required; planned for a later release |
| `DressCode-1024/unet` | DressCode | 1024×768 | Checkpoint required; planned for a later release |

The 1024 checkpoints are not currently included in this repository.

Legacy paths such as `VITON-HD/checkpoint-16000/unet` and
`DressCode/checkpoint-32000/unet` are retained for existing users.

## Installation

```bash
git clone https://github.com/Levinna/DeCo-VTON.git
cd DeCo-VTON
pip install -r requirements.txt
pip install -e .
```

## Usage

> This repository contains UNet checkpoints, not a standalone Diffusers
> pipeline. `DiffusionPipeline.from_pretrained()` cannot load them directly.
> Install DeCo-VTON and use `DeCoVTONPipeline.from_vton_checkpoint()` as shown
> below.

The pipeline loads the VAE from `stabilityai/sd-vae-ft-mse` and the scheduler
configuration from `stable-diffusion-v1-5/stable-diffusion-inpainting`.

```python
import torch
from decovton import DeCoVTONPipeline

pipe = DeCoVTONPipeline.from_vton_checkpoint(
    hf_repo="levinna/DeCo-VTON",
    subfolder="VITON-HD-512/unet",
    torch_dtype=torch.bfloat16,
).to("cuda")
```

See the [GitHub repository](https://github.com/Levinna/DeCo-VTON) for dataset
preparation, inference, and evaluation instructions.

## License

The model weights are licensed under
[CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/), reflecting the
non-commercial terms of the VITON-HD and DressCode datasets. The accompanying
source code is licensed separately under CC BY-NC-SA 4.0.

## Citation

```bibtex
@article{na2025rethinking,
  title={Rethinking Garment Conditioning in Diffusion-based Virtual Try-On: Decouple, Don't Denoise},
  author={Na, Kihyun and Choi, Jinyoung and Kim, Injung},
  journal={arXiv preprint arXiv:2511.18775},
  year={2025}
}
```