levihsu/OOTDiffusion
OOTDiffusion
OOTDiffusion is an AI virtual try-on system designed for controllable image synthesis. It generates images of people wearing specific clothing items by superimposing garments onto human figures for both half-body and full-body compositions.
The project facilitates digital fashion prototyping and virtual clothing fitting by creating garment-to-person overlays. It aims to maintain the original identity of the wearer and the specific details of the clothing during the synthesis process.
The system utilizes a latent diffusion model and conditioning-based image generation to control the output. It employs a cross-attention mechanism and dual-path feature fusion to align visual features from clothing images to the human figure.
Features
- Latent Diffusion Models - Based on a latent diffusion model that performs iterative denoising within a compressed latent space.
- Fashion Visualization - Facilitates digital fashion prototyping by creating visual previews of clothing on human bodies.
- Feature Fusion Architectures - Employs dual-path feature fusion to merge person and clothing representations for structural alignment and texture preservation.
- Cross-Attention Mechanisms - Utilizes cross-attention mechanisms to align garment visual features with specific spatial regions of the human figure.
- Image-Conditioned Generation - Implements generation guided by garment images and human poses as structural and stylistic references.
- Variational Autoencoders - Utilizes a variational autoencoder to compress high-resolution images into low-dimensional tensors.
- Iterative Denoising Pipelines - Uses a U-Net based architecture for the iterative denoising process essential to image synthesis.
- Human Image and Video Generation - Generates realistic images of human figures wearing specific garments for half-body and full-body compositions.
- Virtual Try-On Systems - Provides a deep learning system for mapping clothing items onto human images via virtual try-on.
- Image-to-Image Synthesis Frameworks - Implements a framework for merging garment and person images using reference guidance for precise overlays.
- Virtual Fitting Rooms - Enables virtual fitting by visualizing how specific garments look on human figures for e-commerce applications.
- Virtual Try-On - Uses outfitting fusion for controllable virtual try-on.
