The MAG-Edit method enables localized image editing in complex scenarios. This training-free, inference-stage optimization method optimizes the noise latent feature in diffusion models by maximizing two mask-based cross-attention constraints of the edit token, enhancing local alignment with the desired prompt. It is more effective than existing mask-based inpainting methods and mask-free attention-based methods, achieving both text alignment and structure preservation for localized editing within complex scenarios.
Publication date: 18 Dec 2023
Project Page: https://mag-edit.github.io/
Paper: https://arxiv.org/pdf/2312.11396
