Semantic-Aware Local Image Editing with a Single Mask Operation

We introduce a user-friendly method for controllable image editing, where users simply draw an imprecise mask on the reference image to adaptively transfer its stylistic elements to the target image. Our approach, Adaptive Paste-GAN, is an optimization-based method that relies on intermediate feature maps of GANs for supervision. The method consists of two stages: ROI detection and local editing. In the ROI detection stage, deformable feature matching identifies the optimal editing region within the StyleGAN feature maps. In the editing stage, the latent code is optimized to align the target image's ROIfeatures with those of the reference, while applying regularization to minimize changes outside the ROI. Experimental results demonstrate the precision of ROI detection and show that our method effectively balances locality and global consistency during optimization, and aligns well with user intent across various image categories. The code will be made available upon publication.