ladi-vton
LaDI-VTON is an advanced software implementation for virtual try-on tasks, presented at ACM Multimedia 2023. It integrates latent diffusion models with textual inversion to generate realistic images of models wearing specific in-shop garments. The architecture utilizes a novel autoencoder module with learnable skip connections to preserve the models characteristics while incorporating a textual inversion component that maps garment visual features to CLIP token embeddings. This allows the system to effectively maintain texture and details during generation. The solution has demonstrated superior performance over existing competitors on the Dress Code and VITON-HD benchmarks. Designed for e-commerce and metaverse applications, this tool enables high-fidelity image synthesis by conditioning the generation process with pseudo-word token embeddings derived from input garments. The system relies on PyTorch, diffusers, and transformers libraries and requires a Conda environment for setup. It represents a significan