Virtual Try-On is revolutionizing the way customers shop online. It lets customers experience products virtually before purchasing, boosting their confidence and satisfaction with their online shopping decisions.
In this article, I share the top 6 research papers for those researching Virtual Try-On and seeking AI research and development assistance.
Outfit Anyone: Ultra-high quality virtual try-on for Any Clothing and Any Person
Outfit Anyone is a Virtual Try-On AI model from Humanaigc, crafted for virtual clothing outfitting. The system facilitates ultra-high-quality virtual try-ons for any clothing and any person. It has also been highlighted in discussions on integrating other AI technologies for dressing and animating models…
Diffuse to Choose: Enriching Image Conditioned Inpainting in Latent Diffusion Models for Virtual Try-All
As online shopping expands, the capability for buyers to virtually visualize products in their environments—a concept we term “Virtual Try-All”—has become essential. Recent diffusion models inherently possess a world model, making them suitable for this task within an inpainting context. Yet, traditional image-conditioned diffusion models often struggle to capture the intricate details of products. In contrast, personalization-driven models like DreamPaint excels at retaining the item’s intricacies but are not tailored for real-time applications. We introduce “Diffuse to Choose,” an innovative diffusion-based image-conditioned inpainting model that adeptly balances rapid inference with the retention of high-fidelity details in a reference item while ensuring accurate semantic manipulations in the scene content. Our method relies on incorporating intricate features from the reference image directly into the latent feature maps of the central diffusion model, coupled with a perceptual loss to further conserve the reference item’s details. We perform comprehensive testing on proprietary and public datasets, demonstrating that Diffuse to Choose outperforms existing zero-shot diffusion inpainting methods and few-shot diffusion personalization algorithms like DreamPaint…
TryOnDiffusion: A Tale of Two UNets
Given two images depicting a person and a garment worn by another, our objective is to visualize how the garment might appear on the input person. A significant challenge is synthesizing a photorealistic, detail-preserving visualization of the garment while adapting the garment to fit significant body pose and shape differences between subjects. Previous methods either concentrated on garment detail preservation without effectively accommodating pose and shape variations or facilitated try-ons with the desired shape and pose but compromised on garment details. This paper introduces a diffusion-based architecture that integrates two UNets (Parallel-UNet). This enables us to maintain garment details and adapt the garment for significant pose and body changes within a single framework. The fundamental concepts behind Parallel-UNet include 1) implicit garment warping through a cross-attention mechanism and 2) garment warping and person blending occurring as part of a unified process rather than as a sequence of separate tasks. Experimental outcomes show that TryOnDiffusion delivers state-of-the-art performance both qualitatively and quantitatively…
StableVITON: Mastering Semantic Correspondence with Latent Diffusion Model for Virtual Try-On
Given an article of clothing and a person’s image, the goal of an image-based virtual try-on is to create a tailored image that looks authentic and precisely mirrors the features of the clothing piece. In this study, we endeavor to broaden the use of the pre-trained diffusion model, enabling its independent application for the virtual try-on challenge. The primary obstacle lies in retaining the intricate details of the clothing while effectively leveraging the robust generative power of the pre-trained model. To address these challenges, we introduce StableVITON, a method for mastering the semantic correspondence between clothing and the human figure within the latent domain of the pre-trained diffusion model in an integrated fashion. Our innovative zero cross-attention blocks not only conserve the clothing’s nuances by mastering semantic links but also produce high-quality images by tapping into the intrinsic capabilities of the pre-trained model during the warping phase. Employing our novel attention to total variation loss and applying augmentation, we secure a crisp attention map, enhancing the depiction of clothing nuances. StableVITON surpasses the benchmarks in both qualitative and quantitative assessments, delivering high-quality results for arbitrary person images.
High-Resolution Virtual Try-On Addressing Misalignment and Occlusion Challenges
Image-based virtual try-on aspires to compose an image of an individual donning a specified piece of clothing. Traditional approaches morph the clothing item to suit the individual’s physique and create the segmentation map of the person adorned with the item before amalgamating the item with the person. However, executing the morphing and segmentation creation stages separately without mutual information exchange leads to misalignment between the morphed clothes and the segmentation map, culminating in imperfections in the final image. This disconnection also triggers excessive morphing near the clothing areas concealed by body parts, known as pixel-squeezing flaws. To rectify these issues, we introduce an innovative try-on condition generator that amalgamates the two phases (i.e., morphing and segmentation creation) into a single module. A newly devised feature fusion block within the condition generator facilitates information exchange, eliminating misalignment or pixel-squeezing flaws. Moreover, we implement discriminator rejection to weed out inaccurate segmentation map forecasts, bolstering the efficacy of virtual try-on systems. Tests on a high-resolution dataset reveal that our model manages misalignment and occlusion, significantly advancing beyond the benchmarks.
Harnessing the Potential of Diffusion Models for Premium-Quality Virtual Try-On with Appearance Flow
Virtual try-on is an essential task in image synthesis that focuses on transferring apparel from one image to another while preserving the integrity of the human figure and the garments. Although numerous existing methods utilize Generative Adversarial Networks (GANs) to accomplish this, imperfections often arise, especially at higher resolutions. The diffusion model has been recognized as a formidable alternative for producing premium-quality images across diverse applications. Nonetheless, merely employing clothing to guide the diffusion model’s inpainting process falls short of maintaining the garment’s details. To surmount this hurdle, we propose a model-based inpainting strategy that employs a warping module to steer the diffusion model’s generative process effectively. This warping module preliminarily processes the apparel, preserving the garment’s local details. Subsequently, we merge the processed apparel with a clothes-agnostic human image and introduce noise as the diffusion model’s input. Additionally, the processed apparel serves as local conditions for each denoising step, ensuring the preservation of maximal detail in the output. Our strategy, dubbed Diffusion-based Conditional Inpainting for Virtual Try-ON (DCI-VTON), not only harnesses the diffusion model’s capabilities but also yields realistic and high-fidelity virtual try-on results through the integration of the warping module.
Rienstra' Machine Learning Blog
Machine Learning, Machine Learning Ai, Machine Learning Tools, Machine Learning Course, Machine Learning Engineer