Published Paper Details:

The advancement of diffusion-based generative models has significantly transformed controllable image synthesis, enabling realistic rendering across diverse domains such as editing, and semantic manipulation. Despite these developments, virtual try-on--the task of synthesizing high-fidelity images of a target person wearing a new garment--remains technically challenging due to complex interactions between human pose, garment geometry, occlusions, and appearance fidelity. Existing GAN-based and deformation-based systems often struggle with texture misalignment, identity drift, and loss of garment details in high-resolution scenarios.This paper introduces SDM-VTON, a diffusion-based virtual try-on framework built upon the SDXL latent diffusion model and enhanced with Garment-Guided Cross-Attention (GGCA), a novel conditioning mechanism that integrates spatially aligned garment features into the U-Net denoising process. SDM-VTON includes (1) a multi-scale GarmentNet encoder-decoder that extracts hierarchical clothing features; (2) an image-conditioned adapter that injects CLIP-based semantic embeddings into transformer layers; and (3) a Try-On U-Net augmented to accept pose, segmentation, and mask priors for geometry-aware inpainting. The system is trained end-to-end in latent space, enabling identity preservation, accurate garment transfer, and superior texture reconstruction compared to prior state-of-the-art diffusion models such as LaDI-VTON and FITMI. Extensive experiments on the VITON-HD and Dress-Code datasets demonstrate that IDM-VTON achieves new benchmarks across SSIM, LPIPS, FID, KID, and CLIP Score. This work establishes SDM-VTON as a robust, workspace-native virtual try-on system suitable for next-generation e-commerce, digital dressing rooms, and personalized fashion visualization.