---
license: other
task_categories:
  - image-to-image
language:
  - en
tags:
  - virtual-try-on
  - fashion
  - image-synthesis
  - viton
  - pose-adaptability
  - appearance-flow
pretty_name: VITON-Extends (Train + Test Data)
size_categories:
  - 10K<n<100K
---

# VITON-Extends — Train & Test Data

**VITON-Extends** is an image-based virtual try-on dataset built on [VITON](https://github.com/shadow2496/VITON-HD) with additional challenging poses from [DeepFashion](http://mmlab.ie.cuhk.edu.hk/projects/DeepFashion.html), as described in:

**Enhancing Pose Adaptability in Virtual Try-On Systems**  
Nguyen Dinh Hieu (ORCID:[0009-0002-6683-8036](https://orcid.org/0009-0002-6683-8036))
, Tran Minh Khuong, Phan Duy Hung (ORCID: [0000-0002-6033-6484](https://orcid.org/0000-0002-6033-6484))  
FPT University, Hanoi, Vietnam  

- **Paper (Springer LNCS, IUKM 2025):** [https://doi.org/10.1007/978-981-96-4606-7_21](https://doi.org/10.1007/978-981-96-4606-7_21)  
- **Code repository:** [https://github.com/nguyendinhhieu1309/VITON-Extends](https://github.com/nguyendinhhieu1309/VITON-Extends)  
- **Model weights (Hugging Face):** [NguyenDinhHieu/VITON-Extends](https://huggingface.co/NguyenDinhHieu/VITON-Extends)

## Abstract

Accurate garment fitting in virtual try-on systems remains difficult under complex body poses, occlusions, and large misalignments between person and garment images. The VITON-Extends line of work improves **pose adaptability** and **garment warping** using a **global appearance flow** estimator with **StyleGAN-style** global modulation and a **local flow refinement** stage. Experiments on the VITON benchmark show strong results, especially in challenging poses.

**Keywords:** virtual try-on, pose adaptability, garment warping, StyleGAN, global appearance flow estimation, VITON benchmark.

## Dataset summary

| Property | Value |
|----------|--------|
| Focus | Upper-body virtual try-on (women), multi-pose |
| Typical resolution | 1024 × 768 |
| Splits on Hub | Two archives at repo root: **`Train.zip`** and **`Test.zip`**. **Extract both** after download to obtain the **`train/`** and **`test/`** folder layout below |
| Annotations | Parsing-style labels, edges, DensePose-related assets (see folder layout) |

VITON-Extends extends the original VITON setting with **more diverse standing postures and arm configurations** (e.g. arms crossed, sideways stance), which stress-tests warping and synthesis under occlusion and misalignment.

## Folder layout (this upload)

The Hugging Face dataset repository root ships **`Train.zip`** and **`Test.zip`** (not the raw folders). **You must unzip both** into the same parent directory (e.g. your `local_dir` after `snapshot_download`) so you end up with **`train/`** and **`test/`** as described below. If an archive already contains a single top-level `train` or `test` folder, extract as usual; if your tool nests an extra directory, move the inner `train` / `test` so `dataroot` points at the folder that **directly** contains `train_img/`, `test_img/`, etc.

Inside each folder, subfolder names match what the [VITON-Extends](https://github.com/nguyendinhhieu1309/VITON-Extends) code expects **relative to `--dataroot`**: training reads `train_img`, `train_label`, … from the directory you pass as `dataroot`, so **after extracting**, point **`dataroot` to the `train` folder** (the path that directly contains `train_img/`, …). For testing / inference, point **`dataroot` to the `test` folder** (the path that directly contains `test_img/`, `test_clothes/`, `test_edge/`).

### Under `train/`

| Directory | Role |
|-----------|------|
| `train/train_img/` | Person / scene RGB images used for training |
| `train/train_color/` | Color-aligned representations (dataset-specific preprocessing) |
| `train/train_edge/` | Edge maps for garment / boundary cues |
| `train/train_label/` | Semantic parsing / label maps |
| `train/train_densepose/` | DensePose-related maps aligned with training images |

### Under `test/`

| Directory | Role |
|-----------|------|
| `test/test_img/` | Person or test-scene RGB images |
| `test/test_clothes/` | Garment / clothing images for try-on |
| `test/test_edge/` | Edge maps aligned with the test setup |

File naming follows the conventions expected by the training and testing scripts in the official code repository.

## Intended use

- Training and evaluating **parser-based** and **parser-free** virtual try-on models that expect VITON-style directory layout.  
- Research on **pose-adaptive** try-on and **appearance-flow** warping.

**Not for:** identifying individuals; any deployment that violates privacy or terms of the underlying source datasets.

## How to load (example)

```python
from huggingface_hub import snapshot_download

# Downloads repo root (including Train.zip and Test.zip) into ./VITON-Extends_data
path = snapshot_download(
    repo_id="NguyenDinhHieu/VITON-Extends-DB",
    repo_type="dataset",
    local_dir="./VITON-Extends_data",
)

# Then unzip Train.zip and Test.zip in that folder (Explorer, unzip, 7-Zip, etc.)
# so you get ./VITON-Extends_data/train/... and ./VITON-Extends_data/test/...

# Training: set dataroot to the train folder, e.g. ./VITON-Extends_data/train
# Testing:  set dataroot to the test folder,  e.g. ./VITON-Extends_data/test

# Only the training archive (smaller download):
# snapshot_download(..., allow_patterns=["Train.zip"])

# Only the test archive:
# snapshot_download(..., allow_patterns=["Test.zip"])
```

Use the **`train`** directory as `dataroot` when running training scripts, and the **`test`** directory as `dataroot` when running inference, per [training / testing instructions](https://github.com/nguyendinhhieu1309/VITON-Extends).

## Citation

If you use this dataset, please cite the paper:

```bibtex
@inproceedings{hieu2025vitonextends,
  title     = {Enhancing Pose Adaptability in Virtual Try-On Systems},
  author    = {Hieu, Nguyen Dinh and Khuong, Tran Minh and Hung, Phan Duy},
  booktitle = {Integrated Uncertainty in Knowledge Modelling and Decision Making (IUKM 2025)},
  series    = {Lecture Notes in Computer Science},
  volume    = {15585},
  publisher = {Springer},
  address   = {Singapore},
  year      = {2025},
  doi       = {10.1007/978-981-96-4606-7_21}
}
```

## Acknowledgements

This dataset builds on **VITON** / community virtual try-on resources and **DeepFashion**-sourced challenging poses, following the methodology described in the paper. The implementation builds on ideas from **ClothFlow** and related appearance-flow try-on works; see the [GitHub repository](https://github.com/nguyendinhhieu1309/VITON-Extends) for full acknowledgements.

## Contact

- hieundhe180318@fpt.edu.vn  
- khuongtmhe180089@fpt.edu.vn  
- hungpd2@fe.edu.vn  

---

### Dataset card (Tiếng Việt — tóm tắt)

Trên Hub, dữ liệu chính nằm ở hai file nén gốc **`Train.zip`** và **`Test.zip`**. Sau khi tải về, **bắt buộc giải nén** cả hai để có thư mục **`train/`** và **`test/`**. Trong `train/` gồm các thư mục con `train_img`, `train_color`, `train_edge`, `train_label`, `train_densepose`; trong `test/` gồm `test_img`, `test_clothes`, `test_edge`. Khi chạy code, đặt `dataroot` trỏ vào thư mục **`train`** (huấn luyện) hoặc **`test`** (suy luận). Mô hình: [NguyenDinhHieu/VITON-Extends](https://huggingface.co/NguyenDinhHieu/VITON-Extends). Vui lòng trích dẫn bài báo qua DOI ở trên.