A high-res virtual try-on dataset with more paperwork than pixels

What it does

Dress Code is a dataset of 53,792 garment pairs—107,584 images total—scraped from YOOX NET-A-PORTER catalogs at 1024×768 resolution across upper-body, lower-body, and dress categories. Each pair includes a model photo and a corresponding garment image, plus pre-computed OpenPose keypoints, SCHP human parsing masks, and DensePose UV mappings. The repository provides an empty PyTorch project for loading the data, though the actual training code remains locked behind a firm collaboration agreement.

The interesting bit

Most virtual try-on papers spend half their energy re-deriving pose estimates and segmentation masks; Dress Code bundles all of that auxiliary data upfront, letting researchers skip straight to the generative modeling. The authors also benchmark their own method against seven prior approaches on a low-resolution 256×192 split, claiming top marks in SSIM, FID, and KID—without releasing the model weights.

Key highlights

• 53,792 garment pairs across three categories (upper body, lower body, dresses) at 1024×768
• Pre-extracted annotations: 18-point OpenPose keypoints, skeletons, 18-class SCHP human label maps, and DensePose dense labels
• Low-resolution benchmark table comparing CP-VTON, PFAFN, ACGPN, and others against the authors’ unpublished method
• Empty PyTorch data loader scaffold provided; training and inference code explicitly withheld
• Dataset gated behind institutional email verification, a mandatory physically-signed release agreement, and weekly manual review

Caveats

• No training code, model weights, or inference pipeline is available; only an empty data-loading project is supplied
• Access is heavily restricted: no private companies, non-institutional emails are rejected, and incomplete or unsigned release agreements are silently ignored
• The benchmark results are reported but not reproducible from this repository alone

Verdict

Academic researchers building image-based virtual try-on systems who need a large, annotated, multi-category fashion dataset will find the preprocessing alone worth the application paperwork. Developers hoping to clone and train a model this afternoon will leave empty-handed.