FlexiEdit: Frequency-Aware Latent Refinement for Enhanced Non-Rigid Editing

目前的图像编辑方法主要使用DDIM反演，采用两个分支扩散方法来保留原始图像的属性和布局。然而，这些方法在非刚性编辑方面遇到挑战，涉及改变图像的布局或结构。我们的全面分析表明，DDIM潜在的高频分量对于保留原始图像的关键特征和布局至关重要，但也显著限制了这些方法的应用。针对这一问题，我们介绍了FlexiEdit，通过减少目标编辑区域中的高频分量来优化DDIM潜在，提高对输入文本提示的忠实度。FlexiEdit包括两个关键组成部分：（1）潜在细化，修改DDIM潜在以更好地适应布局调整，（2）通过重新反演增强编辑忠实度，旨在确保编辑更准确地反映输入文本提示。我们的方法在图像编辑方面取得了显著进展，特别是在执行复杂的非刚性编辑方面，通过比较实验展示了其增强的能力。

FlexiEdit: A Latent Refinement Approach to Complex Non-Rigid Image Editing

FlexiEdit enhances fidelity to input text prompts by refining DDIM latent and reducing high-frequency components in targeted editing areas, addressing the limitations of current image editing methods with non-rigid edits.

FlexiEdit comprises two key components: Latent Refinement and Edit Fidelity Enhancement via Re-inversion. The approach represents notable progress in image editing, particularly in performing complex non-rigid edits, showcasing its enhanced capability through comparative experiments.

Related work in this field includes DDIM Inversion, which is the primary method utilized in current image editing, and other approaches such as GAN-based methods and image inpainting techniques.