Garments2look Eval

garments2look-eval

Garments2Look: A Multi-Reference Dataset for High-Fidelity Outfit-Level Virtual Try-On with Clothing and Accessories — Hu et al. (2026) (arXiv:2603.14153, 2026)

What this evaluates

Probes the ability of virtual try-on and image editing models to synthesize high-fidelity, multi-reference outfit images. It evaluates whether models can preserve fine-grained garment details, maintain correct layering orders, and adhere to specific styling techniques while keeping the target person's pose consistent.

Datasets

• Garments2Look — total 80000; splits: test (-1); repo https://github.com/ArtmeScienceLab/Garments2Look
• DressCode-MR — total ?; splits: test (-1)

Metrics

• FID↓ (primary) — range: other
  • Fréchet Inception Distance measures the distance between feature distributions of real and generated images. Lower is better.
• KID↓ — range: other
  • Kernel Inception Distance uses polynomial kernel statistics to compare image feature distributions. Lower is better.
• SSIM↑ — range: [0, 1]
  • Structural Similarity Index measures perceived change in structural information between images. Higher is better.
• LPIPS↓ — range: [0, 1]
  • Learned Perceptual Image Patch Similarity computes perceptual distance using deep network activations. Lower is better.
• Garment accuracy — range: percent
  • Binary accuracy score judged by a Vision-Language Model (VLM) on whether the generated garment matches the reference.
• Layering accuracy — range: percent
  • Binary accuracy score judged by a VLM on whether the generated layering order matches the ground truth.
• Styling accuracy — range: percent
  • Binary accuracy score judged by a VLM on whether the generated styling technique matches the reference.

Input / output format

Input: Target person image (OOTD), one or multiple reference garment images, and optionally text prompts or skeleton/pose maps.

Output: A single synthetic try-on image combining the target person's pose/body with the reference garments.

Scoring recipe

# Pseudo-code for evaluation protocol
generated_images = model.predict(person_img, ref_garments, prompt)
fid_score = compute_fid(real_images, generated_images)
kid_score = compute_kid(real_images, generated_images)
ssim_score = compute_ssim(real_images, generated_images)
lpips_score = compute_lpips(real_images, generated_images)
garment_acc = vlm_judge(generated_images, ref_garments, "garment_match")
layering_acc = vlm_judge(generated_images, gt_layering, "layering_match")
styling_acc = vlm_judge(generated_images, gt_styling, "styling_match")

Common pitfalls

• Pose inconsistency heavily penalizes FID/LPIPS even if garment fidelity is high, as standard VTON metrics are sensitive to pose preservation.
• VLM-based accuracy metrics for layering and styling lack standardized prompts, risking evaluator bias or inconsistent grading.
• Multi-reference input strategy (N Ref vs 2 Ref) significantly impacts results; models often fuse items or drop layers when cardinality exceeds 4.

Evidence (verbatim from paper)

Table 3: Quantitative comparison on Garments2Look test set. We report results on classical VTON metrics and accuracy metrics judged by VLM. Models | FID↓ | KID↓ | SSIM↑ | LPIPS↓ | Garment↑ | Layering↑ | Styling↑

Citation

@misc{hu2026garments2look,
  title={Garments2Look: A Multi-Reference Dataset for High-Fidelity Outfit-Level Virtual Try-On with Clothing and Accessories},
  author={Hu et al. (2026)},
  year={2026},
  note={arXiv:2603.14153}
}

• arXiv: 2603.14153