A systematic review of computer vision and image processing literature reveals that deep learning has transformed virtual try-on technology from early physical capture pipelines into sophisticated neural architectures capable of dynamic clothing deformation and photorealistic synthesis. Published in the Journal on Image and Video Processing, the comprehensive survey analyzes 109 core analytical samples selected through a rigorous four-stage screening protocol across six primary databases: Web of Science, IEEE Xplore, Springer Link, arXiv, Google Scholar, and CNKI. Covering research published between January 1, 2017, and May 31, 2025, the review details how algorithmic design addresses persistent retail bottlenecks, notably the inability to try garments on in person during online shopping.
The study establishes a fundamental categorization of virtual try-on systems based on their underlying data representations, contrasting traditional three-dimensional modeling paradigms against two-dimensional image-based frameworks. Early three-dimensional try-on schemes depended primarily on physical scanning devices, laser sensors, and optical tracking equipment to record high-density human point clouds and surface parameters. While these non-deep learning installations achieved millimeter-level human body measurements suitable for bespoke tailoring and professional garment displays, their exorbitant deployment expenses restricted penetration to less than five percent of commercial settings. Deep learning has introduced neural implicit representations, such as neural radiance fields, along with monocular and multiocular depth estimation models that reconstruct spatial geometry from consumer mobile devices or ordinary RGB cameras. However, the researchers emphasize that three-dimensional pipelines continue to face considerable hurdles, as intensive computational demands and device requirements contrast with the immediate latency expectations of mass e-commerce platforms, leaving two-dimensional image approaches as the practical standard for high-volume retail environments.
Within two-dimensional image-based pipelines, the core technical framework relies heavily on convolutional neural networks to coordinate three foundational stages: human body and garment feature extraction, spatial pose alignment, and final image synthesis. The historical trajectory of two-dimensional try-on has transitioned through three recognizable developmental phases, moving from basic geometric fitting to detailed textural optimization and ultimately toward dynamic controllability through modern generative backbones. Human feature extraction serves as the initial anchor, isolating skeletal structures, joint orientations, and surface boundaries to guide subsequent clothing deformation. In contemporary systems, this structural extraction depends primarily on three standard frameworks: OpenPose, DensePose, and HRNet. Each framework demonstrates distinct operational trade-offs across spatial precision, computational complexity, and inference latency, determining its suitability across varied commercial deployment scenarios.
OpenPose operates as a real-time, deep learning pose estimation framework that identifies eighteen key joint coordinates across single-person or multi-person imagery using convolutional layers and part affinity fields. By modeling associative relationships without requiring manual three-dimensional labels, OpenPose facilitates rapid spatial-semantic alignment on modest hardware, achieving adoption in baseline pipelines such as CP-VTON, CP-VTON+, ACGPN, GP-VTON, MG-VTON, DOC-VTON, and COTTON. Nevertheless, because its outputs remain strictly two-dimensional, OpenPose struggles with joint occlusions and lacks the geometric depth needed to anchor delicate fabric patterns under complex poses. Conversely, DensePose provides pixel-level three-dimensional human surface modeling by mapping individual pixels to spatial and UV coordinate systems, allowing networks like LGVTON, SD-VITON, HR-VITON, and DCI-VTON to perform high-resolution texture mapping. This precision carries substantial computational costs that limit mobile responsiveness. HRNet offers an effective middle ground, preserving high-resolution feature representations across multi-branch sub-networks to achieve superior joint localization for single-person try-on tasks without requiring enterprise graphics hardware.
Complementing human pose estimation, garment feature extraction requires networks to capture both localized textile structures and broad garment silhouettes. The review traces an evolution from local convolutional kernels to global attention sequences. Classic convolutional backbones such as VGGNet employ compact three-by-three kernels that isolate localized textures, such as the grid distribution of checkered shirts, whereas ResNet leverages residual skip connections to train deeper architectures that characterize overarching silhouettes and cut styles, identifying whether an item is loose or form-fitting. Multi-branch Inception v3 networks simultaneously capture intricate ornamental trims and global garments. In style translation workflows, generative models like Pix2Pix and CycleGAN bridge sketch-to-image translations and stylistic conversions. More recently, vision transformers segment garment imagery into linear patch sequences, enabling global self-attention mechanisms to map complex decorative motifs and structural layout relationships across garments that challenge conventional local convolution filters.
Once features are extracted, clothing warping algorithms deform target garments to match the target human posture, addressing spatial misalignments through non-rigid geometric transformations. The review identifies two dominant warping philosophies: parametric transformations based on thin-plate splines and non-parametric displacement fields powered by appearance flow. Thin-plate spline methods, utilized in VITON and CP-VTON, learn geometric matching modules that align paired coordinate points across images via radial basis functions. This regional deformation logic achieves rapid inference speeds of 10 to 15 frames per second, providing structural stability for relaxed human postures. However, thin-plate splines struggle under extreme joint angles, where deformation errors can exceed 8 percent, producing compressed patterns or stretched artifacts. In contrast, appearance flow methods estimate dense, pixelwise vector fields that direct local pixel migration. By calculating flow fields across designated garment sub-segments, frameworks like Clothflow and local appearance networks reduce geometric fitting errors below 3 percent, preserving intricate patterns at the expense of computational overhead and slower frame rates between 5 and 8 frames per second.
The synthesis stage responsible for generating final composite imagery has undergone a major paradigm shift from generative adversarial networks to progressive diffusion models. Generative adversarial networks deploy paired generator and discriminator architectures optimized through zero-sum min-max games, producing composite outputs directly in a single forward pass. Networks such as VITON-HD and HR-VITON integrate misalignment-aware normalization and occlusion-handling modules to render high-resolution 1024-by-768 images. Despite their rapid execution and suitability for real-time virtual fitting, adversarial models struggle with complex multimodal conditioning and frequently exhibit surface blurring, unnatural boundary artifacts, or garment penetration when subjects cross their arms or bend forward. Latent diffusion models overcome these shortcomings by treating image generation as a parameterized reverse denoising process across stochastic noise schedules, mapping inputs into compressed latent spaces to retain fine fabric wrinkles, natural shadow gradients, and delicate edge definitions without prohibitive hardware costs.
Quantitative and qualitative benchmarks compiled in the review underscore the technical superiority of diffusion formulations across paired and unpaired datasets. Evaluating models on the standard VITON-HD benchmark using an NVIDIA GeForce RTX 4090D platform, researchers observed that diffusion frameworks such as LaDI-VTON, StableVITON, and OOTDiffusion systematically outperformed adversarial counterparts. In unpaired high-resolution evaluations, where absent ground-truth imagery precludes structural similarity and learned perceptual patch measurements, diffusion architectures demonstrated superior distributional fidelity, with OOTDiffusion achieving a Frechet Inception Distance score of 8.48 and a Kernel Inception Distance of 0.154, closely followed by LaDI-VTON at 8.57 and 0.153. Across paired and unpaired configurations, diffusion methods improved Frechet Inception Distance metrics by more than 41 percent over traditional generative adversarial models, establishing noticeable advantages in rendering complex textiles, preserving zebra-striped fabric patterns, and generating plausible garment deformations around articulated joints without floating artifacts.
Despite these algorithmic achievements, commercial virtual try-on systems encounter severe operational bottlenecks, physical simulation limitations, and pervasive socio-technical biases. Moreover, hybrid two-dimensional to three-dimensional reconstruction techniques often exhibit fitting discrepancies between 5 and 8 millimeters during severe body articulation, failing the strict two-millimeter tolerance demanded by bespoke tailoring. The review also identifies critical representation disparities across standard benchmark datasets: over 70 percent of samples in VITON-HD and DeepFashion represent standard body mass index ranges between 18.5 and 23.9, resulting in distorted warping predictions and systemic exclusion for high-BMI individuals and non-Western body shapes, while simultaneously raising critical consumer privacy concerns regarding the collection, storage, and potential memorization of sensitive biometric photography.
Looking forward, the authors outline three strategic research pathways necessary to achieve practical, inclusive, and industrially viable virtual try-on ecosystems. First, the discipline must advance lightweight three-dimensional modeling by coupling monocular input imagery with neural implicit representations, democratizing accurate volumetric fitting on standard consumer graphics processing units without expensive physical scanning rigs. Second, architectures must embrace multimodal foundation models, combining contrastive language-image pretraining with reinforcement learning to align natural-language styling prompts, interactive user behavior, and multi-piece outfit layering. Finally, the field requires large-scale, ethically balanced benchmarks documenting diverse human morphologies, dynamic movements, and multi-clothing interactions, supported by occlusion-aware fusion modules. By synthesizing structural precision, progressive denoising, and balanced representative data, future virtual try-on research promises to deliver robust, personalized, and photorealistic fitting experiences for the next generation of global digital commerce.
Subject of Research: Deep learning technologies for virtual try-on systems
Article Title: A review of deep learning-based virtual try-on research
Article References: Fu, X., Geng, S., Pu, W., & Dang, X. (2026). A review of deep learning-based virtual try-on research. Journal on Image and Video Processing, 2026(1), Article 5. https://doi.org/10.1186/s13640-026-00691-w
Image Credits: AI Generated
DOI: 10.1186/s13640-026-00691-w
Keywords: Virtual try-on, Deep learning, Convolutional neural networks, Generative adversarial networks, Diffusion models, Clothing warping, Feature extraction, Thin-plate spline, Appearance flow, DensePose, OpenPose, HRNet, Geometric matching, Image synthesis, VITON-HD, Latent diffusion models, OOTDiffusion, LaDI-VTON
Cite Scienmag News
Blake Davidson. (November 18, 2026). Review Evaluates Deep Learning Techniques Shaping Two Dimensional and Three Dimensional Virtual Fitting. Scienmag. https://scienmag.com/review-evaluates-deep-learning-techniques-shaping-two-dimensional-and-three-dimensional-virtual-fitting/
Blake Davidson. "Review Evaluates Deep Learning Techniques Shaping Two Dimensional and Three Dimensional Virtual Fitting." Scienmag, 18 November 2026, https://scienmag.com/review-evaluates-deep-learning-techniques-shaping-two-dimensional-and-three-dimensional-virtual-fitting/. Accessed 18 November 2026.
Blake Davidson. "Review Evaluates Deep Learning Techniques Shaping Two Dimensional and Three Dimensional Virtual Fitting." Scienmag. November 18, 2026. https://scienmag.com/review-evaluates-deep-learning-techniques-shaping-two-dimensional-and-three-dimensional-virtual-fitting/








