cs.CVMar 14, 2026
SaveGarments2Look: A Multi-Reference Dataset for High-Fidelity Outfit-Level Virtual Try-On with Clothing and Accessories
Organizations: The Hong Kong Polytechnic University · Huhu AI Inc.
Abstract
Virtual try-on (VTON) has advanced single-garment visualization, yet real-world fashion centers on full outfits with multiple garments, accessories, fine-grained categories, layering, and diverse styling, remaining beyond current VTON systems. Existing datasets are category-limited and lack outfit diversity. We introduce Garments2Look, the first large-scale multimodal dataset for outfit-level VTON, comprising 80K many-garments-to-one-look pairs across 40 major categories and 300+ fine-grained subcategories. Each pair includes an outfit with 3-12 reference garment images (Average 4.48), a model image wearing the outfit, and detailed item and try-on textual annotations. To balance authenticity and diversity, we propose a synthesis pipeline. It involves heuristically constructing outfit lists before generating try-on results, with the entire process subjected to strict automated filtering and human validation to ensure data quality. To probe task difficulty, we adapt SOTA VTON methods and general-purpose image editing models to establish baselines. Results show current methods struggle to try on complete outfits seamlessly and to infer correct layering and styling, leading to misalignment and artifacts.
Figures & tables
Figure 1: Comparison of data formats in virtual try-on datasets. Outfit-level dataset is collected and generated from a large-scale set of real images, each paired with diverse clothing and accessories, and including the information of outfit layering and styling.
| Dataset Name | Time | Level | #Class | #Image | #Item | #Pair | #Ref Image/Pair | Look Resolution | Source | Layering | Styling | Publicity |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| VITON-HD [ 5 ] | CVPR 21 | Item | 1 | 27358 | 13679 | 13679 | 1 | 1024 768 | R | ✗ | ✗ | Link |
| DressCode [ 27 ] | ECCV 22 | Item | 3 | 107584 | 53792 | 53792 | 1-2 (Avg 1.08) | 1024 768 | R | ✗ | ✗ | Link |
| M&M VTO [ 58 ] | CVPR 24 | Outfit | 2 | - | - | 18.8M | 2 | 1024 512 | R | ✗ | 5 Types | ✗ |
| Street TryOn [ 9 ] | WACV 25 | Item | 13 | 14453 | 14453 | 0 | 1 | 512 320 | R | ✗ | ✗ | Link |
| Shining Yourself [ 26 ] | CVPR 25 | Item | 4 | - | - | 64000 | 1 | 512 512 | R | ✗ | ✗ | ✗ |
| BootComp [ 6 ] | CVPR 25 | Outfit | 7 | - | - | 54000 | 1-4 | 512 384 | R+S | ✗ | ✗ | ✗ |
Table 1: Comparisons of image VTON datasets. We focus on outfit-level VTON. Our data is built upon real high-solution real-world garment and model images. Look Resolution = The resolution (height width) of target image. / = The number of pixel is less than / approximately equal to the product. R/S = Real/Synthetic data. Layering/Styling = The dataset contains annotation of the garment layering order / styling techniques. VLM = Annotations are generated by visual-language models. Publicity = The dataset is open-sourced.
Figure 2: Overview of Garments2Look construction process.
Figure 3: Data distribution statistics for Garments2Look.
| Method | Paired | Unpaired | ||||
|---|---|---|---|---|---|---|
| FID | KID | SSIM | LPIPS | FID | KID | |
| CatVTON [ 7 ] | 16.131 | 6.980 | 0.856 | 0.106 | 18.339 | 7.458 |
| IP-Adapter [ 52 ] | 14.459 | 4.144 | 0.861 | 0.089 | 24.139 | 10.783 |
| FitDiT [ 17 ] | 14.722 | 5.471 | 0.850 | 0.122 | 15.956 | 5.645 |
| FastFit [ 57 ] | 9.311 | 1.512 | 0.859 | 0.079 | 12.059 | 2.123 |
| Nano-Banana [ 8 ] (N Ref) | 15.220 | 6.703 | 0.801 | 0.199 | 15.369 | 6.916 |
Table 2: Quantitative comparison on DressCode-MR [ 57 ] for multi-reference VTON. We evaluate the performance of Nano Banana (Gemini-2.5-Flash-Image) [ 8 ] by two reference strategies.
Figure 4: Comparison of results of 3 SOTA VTON models and 4 general-purpose image editing models on 4 real representative examples from Garments2Look test set. QIE-2509 = Qwen-Image-Edit-2509, NB = Nano Banana, N Ref = Using a model image and multiple single garment images as input, 2 Ref = Using a model image and an OOTD image as input. A yellow box denotes the difference with the look image (GT). A black arrow indicates a distinct artifact boundary. Row 1: 4 items, 1 layer, no accessory. Row 2: 5 items, 2 layers, no accessory. Row 3: 8 items, 3 layers, 2 accessories. Row 4: 9 items, 3 layers, 3 accessories.
| Models | FID↓ | KID↓ | SSIM↑ | LPIPS↓ | Garment↑ | Layering↑ | Styling↑ |
|---|---|---|---|---|---|---|---|
| VTON Models | |||||||
| IP-Adapter [ 52 ] | 6.5511 | 6.7624 | 0.7960 | 0.1472 | 0.4950 | 0.6036 | 0.5693 |
| BootComp [ 6 ] | 8.6301 | 8.9129 | 0.6110 | 0.3364 | 0.5369 | 0.3127 | 0.3547 |
| OmniTry [ 12 ] | 6.5559 | 10.0715 | 0.7244 | 0.2295 | 0.4612 | 0.1674 | 0.2609 |
| FastFit [ 57 ] | 3.5919 | 4.5827 | 0.8550 | 0.0956 | 0.6237 | 0.1305 | 0.3400 |
| Image Editing Models |
Table 3: Quantitative comparison on Garments2Look test set. We report results on classical VTON metrics and accuracy metrics judged by VLM. QIE-2509 = Qwen-Image-Edit-2509, NB = Nano Banana, NBP = Nano Banana Pro.
Figure 5: Garment consistency with respect to the number of reference garment images, group by different types of model.
| Textual Input Type | FID↓ | KID↓ | SSIM↑ | LPIPS↓ |
|---|---|---|---|---|
| Item Category | 23.272 | 1.271 | 0.817 | 0.141 |
| Item Category + Outfit Overall | 22.135 | 0.752 | 0.810 | 0.155 |
| Item Category + Layering & Styling | 21.831 | 0.733 | 0.814 | 0.148 |
| Item Category + Shape & Pose | 21.783 | 0.750 | 0.823 | 0.133 |
| Item Category + Shape & Pose + Layering & Styling | 21.545 | 0.641 | 0.825 | 0.131 |
Table 4: Study of the detail level of the text prompt on sampled data of Garments2Look test dataset, generated by Nano Banana.
Figure 6: Visualization of the contribution of text modality.
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
| Type | Train | Test | Total |
|---|---|---|---|
| Real | 39819 | 402 | 40222 |
| Synthetic | 39278 | 541 | 39819 |
| Total | 79097 | 944 | 80041 |
Table S1: Division of Garments2Look.
Figure S2 : More layering & styling in Garments2Look.
| Method | KID | SSIM | LPIPS | Garment Acc. | Layering Acc. | Styling Acc. | ||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| IP-Adapter | 8.3390 / | 8.0255 | 0.8588 / | 0.7492 | 0.1076 / | 0.1768 | 0.5224 / | 0.4747 | 0.6424 / | 0.5948 | 0.6001 / | 0.5506 |
| BootComp | 13.5787 / | 14.3927 | 0.7702 / | 0.4925 | 0.2399 / | 0.4083 | 0.6354 / | 0.4636 | 0.3059 / | 0.3142 | 0.4374 / | 0.3048 |
| OmniTry | 9.5100 / | 14.8091 | 0.8036 / | 0.6654 | 0.2011 / | 0.2507 | 0.5292 / | 0.4105 | 0.1371 / | 0.1743 | 0.3576 / | 0.2024 |
| FastFit | 4.6095 / | 6.1260 | 0.8559 / | 0.8544 | 0.0968 / | 0.0948 | 0.7189 / | 0.5527 | 0.1266 / | 0.1313 | 0.4727 / | 0.2599 |
| GPT-4o (2 Ref) | 3.7019 / | 0.8719 | 0.8179 / | 0.7128 | 0.1378 / | 0.1693 | 0.9320 / | 0.8624 | 0.8987 / | 0.8374 | 0.7044 / | 0.6881 |
| Seedream 4.0 (2 Ref) | 12.4870 / | 13.7404 | 0.8126 / | 0.6786 | 0.1740 / | 0.1640 | 0.7805 / | 0.7643 | 0.8513 / | 0.8875 | 0.6286 / | 0.7124 |
Table S2: Separated evaluation results on Garments2Look test set (golden/synthetic).
| Models | FID↓ | KID↓ | SSIM↑ | LPIPS↓ |
| Seedream 4.0 [ 36 ] | 34.294 / 39.810 (+5.516) | 10.446 / 13.310 (+2.864) | 0.757 / 0.729 (-0.028) | 0.335 / 0.349 (+0.014) |
| QIE-2509 [ 47 ] | 29.030 / 32.773 (+3.743) | 7.142 / 9.654 (+2.512) | 0.827 / 0.825 (-0.002) | 0.116 / 0.114 (-0.002) |
| NB [ 8 ] | 21.545 / 21.484 (-0.061) | 0.641 / 0.867 (+0.226) | 0.825 / 0.824 (-0.001) | 0.131 / 0.131 (+0.000) |
| NBP | 24.293 / 24.469 (+0.176) | 2.160 / 2.217 (+0.057) | 0.797 / 0.800 (+0.003) | 0.174 / 0.179 (+0.005) |
Table S3 : Ablation experiment for using the skeleton image (pose condition) as an addition reference on image editing models. The format is “2 Ref / 3 Ref (3 Ref 2 Ref Delta Value)”.
Figure S3 : Qualitative results after fine-tuning QIE-2509 on 10K random sampled data from Garments2Look training set.
Figure S4 : User study on layering accuracy.
| Style Guidance Example: Minimalist Style | |
|---|---|
| Template | # Outfit Guide for Minimalist style This guide summarizes the core principles of Minimalist fashion, emphasizing intentionality, timelessness, and understated elegance through clean lines, neutral palettes, and high-quality materials. ## I. Core Preferences and Prohibitions Category Preference Prohibition Primary Colors White, Black, Grey, Navy, Beige/Camel Bright neons, Loud primary colors (e.g., electric blue), Overly saturated jewel tones Accents Colors Olive Green, Muted Blush, Terracotta, Soft Khaki, Charcoal Fluorescent shades, Clashing bold hues, Iridescent fabrics Patterns Solid, Subtle Stripe (e.g., pinstripe, thin horizontal), Micro-check, Herringbone (subtle), Small geometric Large florals, Busy animal prints, Loud abstract, Polka dots, Ornate damask Style/Fit Clean lines, Tailored, Structured, Slightly oversized (but intentional), A-line, Straight-cut Bodycon, Overly baggy/shapeless, Excessive ruffles/frills, Unstructured asymmetry Textures/Materials Cotton, Linen, Wool (Cashmere, Merino), Silk, Leather (smooth), Tencel/Modal Cheap synthetics, Overly shiny/glossy, Coarse tweed, Heavily distressed denim Accessories Delicate gold/silver jewelry, Structured leather bag, Classic leather belt, Plain sunglasses, Simple watch, Silk scarf Chunky statement necklaces, Heavily embellished bags, Novelty accessories, Excessive layering of jewelry Overall Vibe/Mood Effortless Chic, Serene, Timeless, Polished, Intentional, Calm Trendy, Flashy, Ostentatious, Cluttered, Overly Casual/Sloppy — ## II. Styling Pattern This stylist’s outfits emphasize intentional simplicity, high-quality foundational pieces, and a harmonious balance of form and function to create a polished and enduring aesthetic. ### 1. Classic Outfit Examples (4 examples) Style Structure (Item Components) Keywords Effortless Everyday White oversized button-down shirt + Tailored black wide-leg trousers + Minimalist leather loafers + Delicate gold hoop earrings + Structured leather tote bag Polished, Comfortable, Versatile Sophisticated Professional Camel wool blazer + Cream silk camisole + Navy straight-leg pants + Low-heeled leather pumps + Sleek minimalist watch Refined, Business-Ready, Streamlined Elevated Casual Grey cashmere crewneck sweater + Black A-line midi skirt + White minimalist sneakers + Small crossbody bag + Simple pendant necklace Relaxed, Chic, Understated Modern Urbanit Black mock-neck long-sleeve top + Light wash straight-leg jeans + Ankle boots + Black tailored trench coat + Oversized sunglasses Contemporary, Sleek, Urban ### 2. Outfit Extended Rules Summary - Color Palette : Dominated by neutrals (black, white, grey, beige, navy), often monochromatic or with subtle tonal variations. Accent colors are muted and used sparingly to maintain serenity and avoid visual clutter. - Layering Principle : Strategic and purposeful, focusing on adding warmth, dimension, or textural interest without creating bulk. Layers should align cleanly and contribute to the overall streamlined silhouette, emphasizing clean lines and functionality. - Fit Requirements : Predominantly clean, tailored, and comfortable. Items should skim the body without being restrictive, or be intentionally oversized for a relaxed yet structured feel. Avoid anything ill-fitting, overly tight, or excessively baggy. - Pattern Restriction : Strict preference for solids. If patterns are used, they are extremely subtle, such as thin pinstripes, micro-checks, or very faint textures, ensuring they do not disrupt the minimalist aesthetic but rather add a subtle layer of interest. - Shoe/Bag Principle : Footwear and bags are chosen for their classic design, high-quality materials, and functionality. Styles are clean, unembellished, and contribute to the overall polished and timeless look (e.g., leather loafers, simple pumps, structured totes). - Accessory Requirements : Minimalist accessories are curated for their understated elegance and practical purpose. They are few in number, high in quality, and serve to complement rather than dominate the outfit. Delicate jewelry, classic watches, and plain sunglasses are preferred. - Overall Balance : Achieved through careful consideration of proportion, texture, and color. Outfits strive for visual harmony, avoiding any single element that overwhelms or distracts from the clean lines and intentional simplicity, creating a sense of calm and effortlessness. - Scene Adaptability : Outfits are designed to be highly versatile, easily transitioning between different settings (e.g., office, casual outings, evening events) with minimal adjustments, thanks to their timeless appeal and foundational nature. |
Table S4 : The prompt template of minimalist style guidance.
Figure S5 : More example data in Garments2Look (1/5).
Figure S6 : More example data in Garments2Look (2/5).
Figure S7 : More example data in Garments2Look (3/5).
Figure S8 : More example data in Garments2Look (4/5).
Figure S9 : More example data in Garments2Look (5/5).
Figure S10 : More comparison results on Garments2Look test set (1/5).
Figure S11 : More comparison results on Garments2Look test set (2/5).
Figure S12 : More comparison results on Garments2Look test set (3/5).
Figure S13 : More comparison results on Garments2Look test set (4/5).
Figure S14 : More comparison results on Garments2Look test set (5/5).
More in this topic
Virtual Try-On
Virtual try-on generates images or videos showing people wearing selected garments while preserving appearance, fit, pose, and identity.
See all 55 papers