Reproduce “RegionE: Adaptive Region-Aware Generation for Efficient 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
Pengtao Chen, Xianfang Zeng, Maosen Zhao, Mingzhu Shen, Peng Ye, Bangyin Xiang, Zhibo Wang, Wei Cheng and 2 more
RegionE is a training-free framework that accelerates instruction-based image editing (IIE) models by exploiting spatial and temporal redundancies. It achieves end-to-end speedups of 2-2.5x on state-of-the-art IIE models like Step1X-Edit, FLUX.1 Kontext, and Qwen-Image-Edit, while maintaining high image quality with PSNR values above 30dB and extremely low perceptual differences (LPIPS around 0.05).
Key insights
- Edited regions and unedited regions in IIE tasks exhibit fundamentally different generation trajectories during denoising, allowing for region-specific acceleration strategies.
- Key and Value pairs in Diffusion Transformer (DiT) attention layers, particularly for unedited and instruction tokens, remain highly stable across diffusion timesteps and can be effectively cached.
- Denoising velocities between consecutive timesteps in edited regions are highly similar in direction and exhibit predictable decay in magnitude, enabling adaptive velocity prediction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.