Reproduce “Cora: Correspondence-aware image editing using few step diffusion” 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

Amirhossein Almohammadi, Aryan Mikaeili, Sauradip Nag, Negar Hassanpour, Andrea Tagliasacchi, Ali Mahdavi-Amiri

Simon Fraser University researchers develop Cora, a framework for image editing using four-step diffusion that handles complex structural changes by integrating correspondence-aware noise corrections that align textures between source and target images using DIFT features, spherical linear interpolation (SLERP) for attention keys and values to achieve smoother feature blending compared to linear interpolation, and Hungarian matching for query permutation to control structural preservation, achieving superior user preference rankings (3.29 vs 2.24 for TurboEdit) while maintaining comparable quantitative metrics to multi-step methods and providing intuitive control parameters α and β for balancing content generation versus preservation and structural adherence versus prompt-driven layout changes.

Key insights

  • Pixel misalignment during structural edits invalidates traditional noise correction terms, requiring correspondence-based realignment for artifact-free results
  • SLERP provides smoother feature transitions than linear interpolation for high-dimensional attention features, while content-adaptive weighting prevents artifacts in newly generated regions
  • Query permutation through Hungarian matching enables fine-grained control over structural preservation versus prompt-driven layout changes

Summary by alphaXiv.

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