Source-linked AI summary
Dress Code: High-Resolution Multi-Category Virtual Try-On
Davide Morelli, Matteo Fincato, Marcella Cornia, Federico Landi, Fabio Cesari, Rita Cucchiara
TL;DR
Image-based virtual try-on has been constrained by limited public datasets that underrepresent lower-body and full-body garments. The paper introduces Dress Code and a pixel-level semantic-aware discriminator, reporting stronger qualitative and quantitative performance than baselines and state-of-the-art competitors.
Problem
Public virtual try-on datasets often lack garment variety, paired images, sufficient scale, or high resolution, especially for lower-body and full-body clothes.
Method
The paper introduces the Dress Code dataset and a virtual try-on architecture using garment warping, human parsing estimation, try-on generation, and PSAD.
Results
The proposed approach outperforms competitors and state-of-the-art architectures quantitatively and qualitatively across clothing categories, resolutions, and multi-garment settings.
Takeaways & Limitations
Dress Code broadens publicly available virtual try-on evaluation to multiple garment macro-categories with high-resolution paired images.
Abstract
from arXiv · showhide
Image-based virtual try-on strives to transfer the appearance of a clothing item onto the image of a target person. Prior work focuses mainly on upper-body clothes (e.g. t-shirts, shirts, and tops) and neglects full-body or lower-body items. This shortcoming arises from a main factor: current publicly available datasets for image-based virtual try-on do not account for this variety, thus limiting progress in the field. To address this deficiency, we introduce Dress Code, which contains images of multi-category clothes. Dress Code is more than 3x larger than publicly available datasets for image-based virtual try-on and features high-resolution paired images (1024x768) with front-view, full-body reference models. To generate HD try-on images with high visual quality and rich in details, we propose to learn fine-grained discriminating features. Specifically, we leverage a semantic-aware discriminator that makes predictions at pixel-level instead of image- or patch-level. Extensive experimental evaluation demonstrates that the proposed approach surpasses the baselines and state-of-the-art competitors in terms of visual quality and quantitative results. The Dress Code dataset is publicly available at https://github.com/aimagelab/dress-code.
1 Introduction
Image-based virtual try-on must preserve a person’s body and pose while fitting garments and retaining their texture, but existing datasets and architectures limit coverage of garment categories and image quality. Dress Code addresses these gaps with a larger, high-resolution multi-category dataset and a pixel-level semantic-aware discriminator.
- Virtual try-on must preserve the person’s intrinsic body shape and pose while fitting the garment and maintaining its original texture.
- Public datasets often lack paired model-garment images, contain few samples, and use low-resolution images, limiting research progress.Dress Code provides 1024 × 768 images and more than 50k image pairs, making it more than 3× larger than VITON.
- Existing architectures are not optimized for upper-body, lower-body, and full-body garments because each class must correspond to the relevant body region.For example, trousers must match leg pose, whereas dresses may require matching the pose from shoulders to hips and knees.
- The proposed architecture anchors garments to the appropriate body region, enabling complete try-on with multiple selected garments.
- PSAD predicts real/generated labels and semantic classes at pixel level, learning image semantics that standard discriminators usually neglect.The semantic-aware discriminator is introduced to improve realism and visual quality in generated try-on images.
- Extensive experiments report that the proposed approach outperforms competitors and state-of-the-art architectures quantitatively and qualitatively across resolutions and multi-garment settings.
2 Related Work
Prior virtual try-on research has pursued coarse-to-fine generation, geometric garment alignment, improved generation strategies, and semantic-layout estimation. These directions target alignment, synthesis, occlusion, pose complexity, and visual quality.
- Early image-based virtual try-on uses coarse-to-fine generation, first predicting a coarse dressed-person image and then refining its texture and shape.
- Geometric transformation modules learn thin-plate spline parameters to improve alignment between garments and body shape.
- Teacher-student and student-tutor-teacher approaches improve generation while enabling training with paired and unpaired data.
- Dress Code examples include pose keypoints, dense poses, and human-body segmentation masks alongside image pairs.
- Semantic-layout methods estimate person parsing or segmentation to handle occlusions, complex poses, and sequential-pose synthesis.The cited work includes conditional segmentation masks, spatio-temporal smoothness, and recurrent person generation.
3 Dress Code Dataset
Dress Code is designed around public availability, paired garment-model images, high resolution, and multiple clothing macro-categories. It contains more than 100k images and enriches model pairs with pose and segmentation annotations.
- The dataset’s design desiderata are public research access, paired garment-model images, high resolution, and multiple clothing macro-categories.
- Dress Code is the first publicly available virtual try-on dataset with multiple macro-categories and high-resolution image pairs.
- The dataset contains more than 100k images evenly split between garments and human reference models.
- Images are collected from YOOX NET-A-PORTER fashion catalogs spanning casual clothes and luxury garments.
- Human-model images are partially anonymized by cropping at the nose to remove physiognomic identity information.
- Each reference model includes 18 OpenPose keypoints, DensePose data, and SCHP-generated segmentation masks for body parts and clothing.
4 Virtual Try-On with Pixel-level Semantics
The baseline virtual try-on architecture warps the garment, estimates the person’s semantic layout, and synthesizes the dressed person. A pixel-level semantic-aware discriminator augments adversarial training by classifying each pixel’s semantics and real/generated status.
- Baseline Architecture: The baseline performs garment warping, human parsing estimation, and final try-on synthesis.These stages respectively align the garment, predict the reference person’s semantic map, and generate the dressed-person image.
- Warping Module: Thin-plate spline warping transforms the input garment to match the reference person’s pose and shape.The warping module learns transformation parameters from the garment and person representations, then applies TPS to produce the warped garment.
- Human Parsing Estimation Module: The parsing module predicts a complete human semantic map from the warped garment, pose image, and masked semantic image.Each output pixel contains a probability distribution over 18 semantic classes, and training uses pixel-wise cross-entropy.
- Try-On Module: The try-on module uses a two-branch U-Net to combine garment and reference-person representations while applying the learned TPS transformation in garment features.The branches separately encode the garment and person information; the decoder combines them to generate the final try-on image.
- Training: The generated image is trained with L1, VGG perceptual, and adversarial losses, with λ_adv weighting the adversarial term.The adversarial formulation uses the proposed Pixel-level Parsing-Aware Discriminator.
- Pixel-level Semantic-Aware Discriminator: PSAD predicts each pixel’s semantic class and whether it is real or generated, using an inverse-frequency-weighted (N + 1)-class cross-entropy loss.The extra class represents fake pixels, while class weighting addresses semantic-class imbalance.
5 Experiments
Experiments evaluate Dress Code across standard and high resolutions, compare PSAD with baselines and competitors, and test multi-garment and VITON settings. Results show stronger realism and visual quality, particularly at high resolution.
- Experimental Setup: Dress Code experiments use 48,392 training pairs, 5,400 test pairs, three resolutions, and unpaired test-set garment-model combinations.The evaluated resolutions are 256 × 192, 512 × 384, and 1024 × 768.
- Experimental Setup: Evaluation uses SSIM for similarity and FID, KID, and IS for realism and visual quality.
- Low-Resolution Results and Ablative Analysis: PSAD outperforms competitors on almost all metrics across clothing categories, while discriminator-based models outperform NoDisc and PSAD trails Patch only slightly on SSIM.
- Low-Resolution Results and Ablative Analysis: PSAD better preserves garment colors, textures, and shapes while reducing artifacts and distortions; dense pose helps with dresses but is not consistently better than human keypoints.
- High-Resolution Results: At 1024×768, PSAD achieves the best FID, KID, and IS results against competitors and the Patch baseline.
- Multi-Garment, User Study, and VITON Results: The study also evaluates sequential upper- and lower-body try-on, human preferences, and VITON comparisons with state-of-the-art architectures.The VITON comparison reports that PSAD increases realism and visual quality; the user-study table states that the model is preferred more than 50% of the time.
6 Conclusion
The paper presents Dress Code as a larger, high-resolution, multi-category virtual try-on dataset and introduces PSAD for improving generated-image realism. Its benchmark covers multiple approaches and supports high-quality try-on across broader clothing categories.
- Dress Code is more than 3× larger than the most common virtual try-on dataset and is the first public dataset with multiple macro-categories and high-resolution images.
- PSAD improves the generation of high-quality images and the realism of virtual try-on results.
- The paper provides a comprehensive benchmark covering up to nine state-of-the-art virtual try-on approaches and different baselines.
Supplementary Material
The supplementary material details Dress Code’s category-specific splits, preprocessing and network components, then reports low- and high-resolution qualitative and quantitative evaluations.
- Dataset: Dress Code reports train/test pairs separately for upper-body, lower-body, and dress categories, as well as for the complete dataset.The experiments test each category separately and the entire test set.
- Preprocessing: Pose representations use OpenPose keypoint heatmaps or concatenated DensePose label and UV maps.OpenPose produces an 18-channel heatmap, while DensePose supplies 25 label-map channels and 2 UV-map channels.
- Preprocessing: The masked person representation removes target-clothing and involved-body information, retaining non-target appearance and body regions for try-on generation.The mask is formed from clothing and limb areas derived from target label and pose maps.
- Network Architecture: The baseline architecture uses separate warping, human-parsing estimation, and try-on modules, with additional U-Net capacity for high-resolution training.The try-on module adds one U-Net block to both encoder and decoder at high resolution; the parsing module also uses a U-Net design.
- Discriminators: PSAD classifies each pixel as a human-parser semantic class or fake, whereas the PatchGAN baseline classifies square patches as real or fake.PSAD uses a U-Net architecture with a final 1 × 1 convolution producing N + 1 output channels.
- Results: Qualitative evaluations cover upper-body clothes, lower-body clothes, dresses, failure cases, PSAD-versus-PatchGAN comparisons, and high-resolution outputs.Table 8 reports complete high-resolution try-on performance on the Dress Code test set, while Figure 16 presents qualitative high-resolution results.