VTBench: Virtual Try-On Benchmark

VTBench is a benchmark dataset for evaluating virtual try-on (VTON) models, introduced in the NeurIPS 2026 Datasets & Benchmarks Track submission.

Test Sets

Subset	Directory	Pairs	Description
FTF (Font/Texture Fidelity)	benchmark_texture/	600	Fine-grained garment texture and text/logo preservation
CBC (Complex Background Consistency)	benchmark_complex_bg/	433	Background preservation under complex scenes
HOC (Hand-Occlusion Consistency)	benchmark_hand/	1,443	Hand-garment occlusion and joint consistency
CCP (Cross-Category Plausibility)	benchmark_crossvton/	420 (4 directions × 105)	Cross-category try-on: dress / upper / lower

Directory Structure

benchmark_texture/
  cloth/              # garment images: cloth_0001.png … cloth_0600.png
  image/              # person images:  image_0001.png … image_0600.png
  test_pairs.txt      # <cloth> <image> per line (600 pairs)
  filename_mapping.txt  # new filename → original filename

benchmark_complex_bg/
  cloth/              # 0000.png … 0473.png
  image/              # 0000.png … 0473.png
  test_pairs.txt      # 433 pairs

benchmark_hand/
  cloth/              # 0000.png … 1442.png
  image/              # 0000.png … 1442.png
  test_pairs.txt      # 1443 pairs

benchmark_crossvton/
  dress/cloth/        # dress garments
  dress/image/        # dress person images
  upper/cloth/        # upper-body garments
  upper/image/        # upper-body person images
  lower/cloth/        # lower-body garments
  lower/image/        # lower-body person images
  test_pairs_ds2lower.txt   # dress → lower (105 pairs)
  test_pairs_ds2up.txt      # dress → upper (105 pairs)
  test_pairs_lower2ds.txt   # lower → dress (105 pairs)
  test_pairs_up2ds.txt      # upper → dress (105 pairs)

Each test_pairs*.txt contains one pair per line: <cloth_path> <image_path>.

Model Outputs

VT_model_results_retinaface_100/ provides a standardized qualitative comparison set for 19 virtual try-on models on 100 identical person-garment input pairs, for a total of 1,900 generated images.

The 100 inputs are the first 100 pairs in benchmark_complex_bg/test_pairs.txt. Every model is evaluated on the same person image and garment image. Faces in the generated results were blurred for privacy. Aggregate preprocessing statistics are reported in processing_report.json.

Evaluated models: CAT-DM, CatVTON, CrossVTON, DA-FLOW, FitDit, FS-VTON, HR-VITON, IDM-VTON, LaDI-VTON, Leffa, Nano-Banana, OOTD, PF-AFN, Qwen-Edit, SD-VITON, Seedream, StableVITON, TPD, and VTON-HandFit.

VT_model_results_retinaface_100/
  results/
    <model_name>/       # 100 generated images per model
      <sample_id>.png
  pairs.tsv             # sample IDs and repository-relative input paths
  model_sources.tsv     # model names and repository-relative result directories
  processing_report.json
  errors.tsv            # processing errors (none for this release)

Each result filename matches the person-image filename for its corresponding row in pairs.tsv. All paths in the manifests are relative to the root of this Hugging Face dataset repository. These generated results are distributed for non-commercial academic use under the same restrictions as the source CBC subset.

License

This dataset is available for non-commercial academic research, teaching, and publication only.

benchmark_crossvton

The benchmark_crossvton/ subset is a curated selection from the Dress Code Dataset (images selected and renamed; no content modification). It is subject to the original Dress Code license terms:

The Dress Code Dataset is proprietary to and © Yoox Net-a-Porter Group S.p.A., and its licensors. It is distributed by the University of Modena and Reggio Emilia, and available for non-commercial academic use under licence terms set out at https://github.com/aimagelab/dress-code.

By downloading or using benchmark_crossvton/, you agree to be bound by those licence terms.

benchmark_texture / benchmark_complex_bg / benchmark_hand

These subsets are released under CC BY-NC 4.0 and are available for non-commercial academic use only.

Citation

Citation will be updated upon publication.