Reproduce “Lazy Diffusion Transformer for Interactive Image Editing” 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

Yotam Nitzan, Zongze Wu, Richard Zhang, Eli Shechtman, Daniel Cohen-Or, Taesung Park, Michaël Gharbi

LazyDiffusion accelerates diffusion-based image editing by encoding the full canvas once for global context, then running iterative denoising only on tokens inside the user-specified mask. For masks covering 10% of a 1024×1024 image, it is about 10× faster than full-image regeneration while achieving similar CLIP alignment, a 7.70 FID, and strong user preference over crop-based baselines.

Key insights

  • Separating global context extraction from repeated local synthesis makes diffusion cost depend primarily on the edited area rather than the full canvas.
  • Retaining context tokens for patches that overlap the mask preserves both scene-level information and boundary details without decoding the entire image.
  • The approach is complementary to faster samplers and distillation: it reduces tokens processed per denoising step rather than reducing the number of steps.

Summary by alphaXiv.

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