OpenVTON

OpenVTON is the dataset released with OpenVTON-Bench: A Large-Scale High-Resolution Benchmark for Controllable Virtual Try-On Evaluation. It contains 99,925 garment-person pairs, with masking images and garment captions, for controllable virtual try-on research and evaluation.

• Paper
• Evaluation code and benchmark documentation
• Croissant metadata with Responsible AI information

Dataset structure

Each row contains the following five fields:

Field	Type	Description
source	Image	Reference garment image
mask	Image	Person image with the target garment region blacked out; masked-person conditioning input
target	Image	Person wearing the garment; target image for the paired evaluation
caption	String	Garment description generated with Gemini 2.0 Flash
category	String	One of 14 coarse labels extracted by keyword matching over captions

The category column contains caption-derived metadata labels. It does not store the DINOv3 visual cluster assignment.

Splits

Split	Pairs
Train	89,927
Validation	4,989
Test	5,009
Total	99,925

The release consists of 18 training, one validation, and two test Parquet shards. Split sizes and statistics here describe the complete release, rather than a dataset-viewer preview.

Visual sampling clusters and caption labels

Final sampling uses 20 visual clusters (C0-C19) obtained by applying K-Means with K=20 to DINOv3 garment embeddings. These are visual sampling groups, not 20 manually defined garment categories.

The separate category field contains 14 coarse labels obtained by deterministic keyword matching over Gemini captions, with first-match precedence and the merges tshirt -> top, knitwear -> sweater, and suit -> jacket:

blazer, cardigan, coat, dress, hoodie, jacket, jeans, pants, shirt, shorts, skirt, sweater, top, and vest.

One caption label can span several visual clusters. Approximately balanced sampling across clusters therefore does not imply uniform caption-label frequencies or demographic balance. The labels are not a manually verified garment taxonomy.

Nineteen clusters contain 5,012-5,088 pairs each. C13 contains all 3,340 available pairs from its source cluster. Sampling is approximately balanced across clusters, subject to source availability, and the splits preserve these cluster proportions.

The project README provides the complete 14 x 20 label-by-cluster cross-tabulation, C0-C19 counts by split, and dataset taxonomy comparison. For example, the full-release frequencies of pants and shorts are 12.01% and 8.97%, respectively.

Usage

from datasets import load_dataset

dataset = load_dataset("RenxingIntelligence/OpenVTON")
sample = dataset["train"][0]
sample["source"].show()  # reference garment
sample["mask"].show()    # person image with the garment region blacked out
sample["target"].show()  # target person image
print(sample["caption"])
print(sample["category"])

To use the data revision described in the accompanying Croissant metadata:

dataset = load_dataset(
    "RenxingIntelligence/OpenVTON",
    revision="afe4443ab7961cabe00ec3122809c89111039cde",
)

Construction and provenance

The dataset combines curated garment-person pairs from the IG Pair dataset associated with IMAGDressing-v1 and publicly available e-commerce imagery. Quality and resolution filtering, including aesthetic scoring and face detection, were followed by human checks of garment-person correspondence and image completeness.

DINOv3 garment embeddings and K-Means define the visual sampling clusters. Gemini 2.0 Flash generates dense garment descriptions using upper-body/lower-body routing and structured prompts. GroundingDINO and SAM3 are used for garment localization and masking; Qwen-VL-Plus supplies post-hoc descriptions for mask prompting. These descriptions do not define cluster membership. The published prompts and paper Section 3 / Appendix A document the construction pipeline.

Synthetic and transformed components: the release combines photographs with generated captions and algorithmically transformed images, including masking and generative facial anonymization. It is not a wholly synthetic-photo dataset. Accordingly, the Croissant metadata declares rai:hasSyntheticData: true and describes the processing activities.

Intended use, limitations, and responsible use

OpenVTON supports non-commercial research on controllable virtual try-on, including model evaluation and adaptation. The accompanying benchmark combines semantic, representation, and pixel-level metrics. Its human-study results concern model-ranking agreement under a shared visual-quality rubric. The benchmark does not establish physical fit, sizing, garment grading, or material physics.

E-commerce and social-media source selection, filtering, and automated annotation may introduce selection or label biases. Extreme poses and complex multi-layer occlusions may be underrepresented. Captions and masks can contain errors, and keyword-matched labels may reflect incidental garment mentions. Demographic representativeness and parity have not been established; the released schema has no demographic annotation fields.

Images depict people and may convey apparent personal attributes. Facial regions underwent generative anonymization before release, as described in the paper. Users should not use the dataset for biometric identification, face recognition, surveillance, or attempts to recover personal identity. Potential misuse also includes deceptive imagery, unauthorized person-image editing, and unrealistic body portrayals. The paper describes a takedown mechanism for individuals and rights holders.

Benchmark configurations

The baseline configurations document the reported model identifiers, checkpoints, inference settings, sampling steps, guidance scales, and seeds. The benchmark applies common aspect-ratio-preserving resizing and centered white padding for evaluation, independently of native inference dimensions. API-based results refer to fixed generated outputs.

Croissant metadata

Use the checked-in croissant.json for the camera-ready metadata submission. It includes both core Croissant fields and the Responsible AI fields requested by NeurIPS 2026, including data provenance, intended uses, limitations, biases, personal-information considerations, social impact, and synthetic components.

The metadata describes all 21 original Parquet shards through direct, revision-pinned download URLs and the SHA-256 checksums published by Hugging Face. Each record set represents one physical shard; concatenate record sets sharing a split prefix to obtain the corresponding train, validation, or test split. The metadata's version 1.0.0+hf.afe4443ab7961cabe00ec3122809c89111039cde identifies this documented data snapshot; it is not a claim that a tag with that name exists on Hugging Face.

Hugging Face also offers automatically generated Croissant metadata. The checked-in file adds dataset-specific RAI information and should be updated whenever data shards change.

License

This dataset is released under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). Sharing and adaptation for non-commercial purposes are subject to attribution and the other license conditions. Original asset terms and third-party rights continue to apply. Recommendations for responsible use are separate from the license terms.

The benchmark code has a separate license documented in its repository.

Citation

If you use this dataset or benchmark, please cite the paper:

@misc{li2026openvtonbenchlargescalehighresolutionbenchmark,
  title={OpenVTON-Bench: A Large-Scale High-Resolution Benchmark for Controllable Virtual Try-On Evaluation},
  author={Jin Li and Tao Chen and Shuai Jiang and Weijie Wang and Jingwen Luo and Chenhui Wu},
  year={2026},
  eprint={2601.22725},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2601.22725}
}