Excavating Consistency Across Editing Steps for Effective Multi-Step Image Editing

Chunyu Qi, Cheng Gu, Zhihui Fu, Chang Jianglong, Jun Wang, Xiaoyao Liang, Haibing Guan, Zhuoran Song

Keywords: region consistency;multi-step image editing;diffusion;acceleration

Abstract: Multi-step image editing with diffusion models typically requires repeatedly executing the inversion–denoising paradigm, which leads to severe challenges in both image quality and computational efficiency. Repeated inversion introduces errors that accumulate across editing steps, degrading image quality, while regeneration of unchanged background regions incurs substantial computational overhead. In this paper, we present ExCave, a training-free multi-step editing framework that improves both image quality and computational efficiency by excavating consistency across editing steps. ExCave introduces an inversion sharing mechanism that performs inversion once and reuses its consistent features across subsequent edits, thereby significantly reducing errors. To eliminate redundant computation, we propose the CacheDiff method that regenerates only the edited regions while reusing consistent features from unchanged background regions. Finally, we design GPU-oriented optimizations to translate theoretical gains into practical reductions in end-to-end latency. Extensive experiments demonstrate that ExCave achieves superior image quality and dramatically reduces inference latency, establishing a new paradigm for accurate and efficient multi-step editing.

Primary Area: generative models

Submission Number: 8300