← Back to Computer Vision
cs.CV

Editing images without destroying what you want to keep

Liangsi Lu, Minzhe Guo, Xuhang Chen, Yang Shi

May 20, 2026

Image editing with diffusion models forces a painful choice: push harder for the semantic change you want, and the image's structure falls apart. The culprit is that edit strength is locked to how noisy the intermediate states get. NaviEdit decouples these by reallocating computation toward the scales where semantic changes actually happen, skipping the wasteful high-noise destruction phase. It's training-free, works on top of existing editors, and shows consistent gains across different models.
Published as Semantic Granularity Navigation in Image Editing arXiv:2605.21190
Read the original paper →