---
title: SIFT-VTON
emoji: 👕
colorFrom: indigo
colorTo: pink
sdk: gradio
sdk_version: 6.24.0
app_file: app.py
short_description: Virtual try-on with SIFT correspondence supervision
python_version: "3.10"
startup_duration_timeout: 1h
license: cc-by-nc-sa-4.0
---

# SIFT-VTON

Interactive demo of **SIFT-VTON: Geometric Correspondence Supervision on Cross-Attention for
Virtual Try-On** (ICPR 2026).

- Paper: https://huggingface.co/papers/2605.01296
- Model: https://huggingface.co/takesuke/SIFT-VTON
- Code: https://github.com/takesukeDS/SIFT-VTON

The sampling path mirrors the authors' `inference_hf.py` (PLMS, 50 steps, cfg 1.5,
`--start_from_noised_agn`, `--repaint`) at the training resolution of 512×384. Because the model
consumes VITON-HD-style inputs (agnostic person image + agnostic mask + DensePose), the Space runs
that preprocessing chain on the fly — SCHP human parsing, OpenPose body keypoints and
DensePose (`densepose_rcnn_R_50_FPN_s1x`) — so arbitrary photographs can be used.

## Credits

The preprocessing code (vendored `detectron2` / `densepose`, human parsing, OpenPose, the VITON-HD
agnostic-mask builder) and the preprocessing checkpoints come from
[IDM-VTON](https://huggingface.co/spaces/yisol/IDM-VTON) (CC BY-NC-SA 4.0). The example person and
garment images are VITON-HD test-set items redistributed from the same Space under CC BY-NC-SA 4.0.

The SIFT-VTON weights are licensed CC BY-NC-SA 4.0 — this demo is for non-commercial research use.