A Review on Dress Sketch/Image Representation from 2D to 3D for Fashion Designing: Using AI

Abstract

This review investigates the growing research on automated conversion from two-dimensional (2D) to three-dimensional (3D) garments and focuses on summarizing the disjointed existing literature in the area of sketch-based 2D-to-3D garment conversion from a fashion sketch and/or a dress image. The review uses a narrative synthesis method and considers seventeen peer-reviewed and preprint studies from 2006 to 2026 that employ artificial intelligence and computational geometry to solve this problem. The key technologies the authors present include Generative Adversarial Networks (GANs) and Convolutional Neural Networks (CNNs) for feature extraction and shape estimation; Graph Long Short-Term Memory (GraphLSTM) networks for recursive feature diffusion across mesh vertices; and the Double Reverse Diffusion (DReD) framework, which integrates multi-scale image features with the unsigned-distance-field autoencoder, implemented in PyTorch and PyTorch3D. In addition, the review also discusses geometric and hybrid approaches, such as selective regularization for curve network reconstruction, unsigned-distance-field based sketch-to-mesh systems and pattern-making techniques which parameterize human body shape to produce flattenable 3D garment structures. Main findings show that existing tools excel in narrow tasks, like shape inference from a single sketch, 3D mesh generation from a single image, or reconstruction of a single garment from an image, but none of them combine the three steps of sketch interpretation, 3D mesh creation and pattern-ready output in a single end-to-end pipeline. Finally, this review suggests that future unified 2D-to-3D fashion design tools should integrate deep-learning-based shape inference, geometric pattern-making, and associative sketch-pattern linkage, and highlights the key challenges of this approach as dataset scarcity, occlusion handling, and estimation of fabric properties.