Reproduce “High-Resolution Virtual Try-On with Misalignment and Occlusion-Handled Conditions” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, Jaegul Choo

HR-VITON, a high-resolution virtual try-on model developed by KAIST and Soongsil University, integrates clothing warping and segmentation generation to resolve pervasive misalignment and pixel-squeezing artifacts. This approach achieves enhanced photo-realism and detail preservation, yielding an FID of 10.91 and KID of 0.179 on 1024x768 images, outperforming previous methods like VITON-HD.

Key insights

  • Integrating clothing warping and segmentation generation within a unified module with explicit information exchange is crucial for eliminating misalignment and occlusion artifacts in high-resolution virtual try-on.
  • Explicit segmentation maps, rather than "parser-free" methods, are vital for high-quality, high-resolution virtual try-on, allowing precise disentanglement of elements and controlled image generation.
  • Multi-scale losses applied to the appearance flow are essential for learning effective coarse-to-fine deformation, and a discriminator rejection mechanism enhances the practical robustness of the system.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.