MAG-Edit: Localized Image Editing in Complex Scenarios via Mask-Based Attention-Adjusted Guidance

• DOI: 10.48550/arxiv.2312.11396
• Publication Date: 2023-01-01
• Authors: Qi Mao; Lan Chen; Yuchao Gu; Zhen Fang; Mike Zheng Shou
• Keywords: Image editing; Inpainting; Feature (linguistics); Video editing

Recent diffusion-based image editing approaches have exhibited impressive capabilities in images with simple compositions. However, localized complex scenarios has not been well-studied the literature, despite its growing real-world demands. Existing mask-based inpainting methods fall short of retaining underlying structure within edit region. Meanwhile, mask-free attention-based often exhibit leakage and misalignment more In this work, we develop MAG-Edit, a training-free, inference-stage optimization method, which enables scenarios. particular, MAG-Edit optimizes noise latent feature diffusion models by maximizing two cross-attention constraints token, turn gradually enhances local alignment desired prompt. Extensive quantitative qualitative experiments demonstrate effectiveness our method achieving both text preservation for