Reproduce “SpecEdit: Training-Free Acceleration for Diffusion based Image Editing via Semantic Locking” 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

Zhengan Yan, Shikang Zheng, Haoran Qin, Xiaobing Tu, Yinggui Wang, Jiacheng Liu, Jiaxuan Ren, Yuqi Lin and 4 more

SpecEdit, a training-free framework, accelerates diffusion-based image editing by dynamically allocating high-resolution computation based on semantic relevance derived from a low-resolution draft. It achieved up to 10x native speedup and 13x combined speedup, while consistently preserving strong semantic consistency and perceptual quality across benchmarks.

Key insights

  • Image edits are inherently sparse; uniform high-resolution processing across the entire image is largely redundant.
  • A low-cost, low-resolution 'draft' inference can effectively provide a semantic prior to accurately localize regions relevant to an edit, guiding dynamic resolution.
  • Combining semantics-aware token selection with sparse global coverage in a dynamic-resolution framework ensures both efficient computation in edited regions and robust structural stability in unedited areas.

Summary by alphaXiv.

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