Reproduce “Diffusion-based Holistic Texture Rectification and Synthesis” 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

Guoqing Hao, Satoshi Iizuka, Kensho Hara, Edgar Simo-Serra, Hirokatsu Kataoka, Kazuhiro Fukui

A framework is presented for automatically generating high-quality, rectified textures from real-world images that contain both occlusions and geometric deformations, utilizing a Latent Diffusion Model with an occlusion-aware latent transformer. The system outperforms existing methods on quantitative metrics and perceptual studies, producing rectified textures from degraded inputs.

Key insights

  • Latent Diffusion Models can effectively generate high-quality, rectified textures when properly conditioned on degraded inputs.
  • A specialized occlusion-aware latent transformer, utilizing partial convolutions and self-attention, is crucial for extracting reliable features from inputs with both occlusions and geometric distortions.
  • Combining channel-wise concatenation and cross-attention provides robust conditioning for the LDM, integrating both global context and fine-grained local guidance.

Summary by alphaXiv.

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