DressCode-MR Virtual Try-On Dataset

• DressCode-MR is a multi-reference dataset offering high-resolution full-body images paired with canonical garment and accessory shots.
• It supports multi-item compositionality by pairing products across five key fashion categories and overcoming occlusion issues via model-based inversion.
• The dataset employs expert networks and human-in-the-loop validation to ensure precise item recovery, quality alignment, and robust style diversity.

DressCode-MR is a large-scale, multi-reference, high-resolution dataset for virtual try-on and related person–apparel pairing tasks. Designed to overcome the limitations of existing datasets—namely, restricted garment categories, absence of accessories, and lack of multi-item compositionality—DressCode-MR provides 28,179 curated sets of paired, full-body images spanning five key fashion categories (tops, bottoms, dresses, shoes, bags). The dataset advances the boundary of virtual try-on research by enabling, for the first time, holistic multi-garment and accessory-aware generation and modeling, with a strict focus on image quality, category alignment, and canonical item recovery. Its construction and usage are detailed below.

1. Dataset Composition and Multi-Reference Structure

DressCode-MR uniquely supports multi-reference pairing by associating each sample with:

• A high-resolution (1024 × 768 px) full-body photograph of a real person in a natural pose and background.
• Up to five product-style, tightly cropped canonical images (one per possible category: top, bottom, dress, shoes, bag), each depicting the actual garment/accessory worn in the person image.
• Outfits may contain any non-overlapping subset (e.g., top+bottom+shoes, dress+bag, etc.), allowing heterogeneous category mixing.

Compared to its predecessors (e.g., Dress Code (Morelli et al., 2022), VITON-HD), DressCode-MR extends coverage to accessories (shoes, bags) in addition to garments, with each item precisely anchored to the instance worn by the modeled subject (Chong et al., 28 Aug 2025).

2. Data Acquisition and Curation Pipeline

The dataset construction leverages:

• The person image pool from the original DressCode dataset (53K person–item pairs) as base material.
• Category-wise expert “item-restoration” networks, each derived from CatVTON/FLUX backbones, trained to invert the appearance of each clothing/accessory from the composite look into a canonical product image. Training incorporates VITON-HD, DressCode, and a small web-scraped aligned shoe/bag set.
• Each DressCode person image is passed through the relevant expert networks, which “extract” the worn item from scene context/occlusion into a stand-alone item shot.
• Human-in-the-loop filtering: Trained annotators rigorously inspect auto-recovered items, excising instances with misalignment, occlusion, or poor reconstruction. The outcome is a final, high-integrity pool of 28,179 samples, each with full and consistent reference alignment (Chong et al., 28 Aug 2025).

3. Statistical Breakdown and Category Distribution

From a total of 28,179 samples:

• Train split: 25,779 samples.
• Test split: 2,400 samples (no pre-defined validation, but subsets may be carved off).

Each sample contains at least one category; combinations are distributed as follows (test set): ~85% contain footwear, ~65% include bags, ~40% contain dresses (the remainder are top+bottom, possibly with accessories). Category frequencies inherit the DressCode garment distribution, augmented by the new accessory classes (Chong et al., 28 Aug 2025).

4. Data Organization, Annotation Formats, and Auxiliary Features

The file structure is strictly folder-based:

/split/{sample_id}/person.jpg /split/{sample_id}/top.jpg /split/{sample_id}/bottom.jpg /split/{sample_id}/dress.jpg /split/{sample_id}/shoes.jpg /split/{sample_id}/bag.jpg

A central CSV/JSON lists each sample_id along with boolean presence flags for each category. Optional auxiliary annotations—cloth-agnostic person mask (AutoMask), 2D pose skeleton (DWPose)—are included or are easily recomputable for conditioning in try-on models and geometric tasks. Images for all modalities are fixed at 1024 × 768 px, facilitating alignment and comparability with benchmarks (Chong et al., 28 Aug 2025).

5. Distinctive Properties and Design Innovations

DressCode-MR is characterized by:

• Canonical item shots recovered via model-based inversion, enforcing consistency in lighting, centering, and deformation across categories—a significant improvement over heterogeneous product images.
• Explicit inclusion of shoes and bags, absent from prior datasets, enabling modeling of complete, realistic outfit compositions including accessories, not only garments.
• Full-body, in-scene person images instead of cropped torsos, promoting robustness in pose, scale, and occlusion handling.
• Accurately paired person–item relationships, where every reference image corresponds to the actual item/instance worn (Chong et al., 28 Aug 2025).

The consistent, automated extraction pipeline coupled with manual curation ensures data quality and cross-category alignment, supporting downstream architectures in handling compositionality and inter-item occlusions.

6. Quality Control, Diversity, and Evaluation

Dataset quality is maintained primarily via a human-in-the-loop filtering protocol; no automatic diversity metrics (e.g., FID, entropy) are reported for the raw dataset. Over 90% of generated item references pass a blind quality check, attesting to high fidelity. Style, color, and pattern diversity reflects the breadth of the DressCode source, encompassing a wide spectrum of in-the-wild fashion. Potential diversity quantification methods include:

• Category co-occurrence analysis (histograms of item pairings).
• Feature-space clustering using pretrained fashion encoders to ensure robust style modality coverage.

No explicit coverage gaps in mainstream apparel or accessories are noted, though detailed attribute-class balance is not quantified (Chong et al., 28 Aug 2025).

7. Research Applications and Benchmarking Utility

DressCode-MR is designed for:

• Training and evaluating multi-reference virtual try-on systems requiring holistic, multi-item compositionality.
• Accessory-aware human image generation and garment-to-person transfer tasks.
• Research on data-efficient item inversion/restoration and style-compatible outfit composition.
• Studying garment co-occurrence patterns and style compatibility for outfit recommendation/analysis.

Its high-resolution, canonical reference design and natural scene context make it a unique resource for pushing the state of the art in realistic and controllable virtual try-on, compositional image synthesis, and robust fashion understanding (Chong et al., 28 Aug 2025).

---

In summary, DressCode-MR constitutes the de facto benchmark for complex, high-resolution, multi-item virtual try-on, distinguished by systematic multi-category and accessory pairing, model-driven canonical recovery, and careful human validation. Its structure and scale respond directly to key bottlenecks in compositional garment modeling, setting a new standard for data-driven research in vision-based fashion applications.