Core Concepts
Optimizing the η function in diffusion inversion enhances real image editing performance.
Abstract
Diffusion models have revolutionized text-guided image editing, but existing methods struggle with faithful edits. The proposed Eta Inversion technique introduces a time- and region-dependent η function to improve editability. By balancing high-level and low-level features, it allows for precise and varied image editing. Through quantitative and qualitative assessments, Eta Inversion outperforms existing strategies, setting a new benchmark in the field. The method not only maintains structural integrity but also significantly improves editing results compared to previous techniques.
Stats
Fig. 1: Existing methods fail to change torch into flower; Eta Inversion creates various plausible results.
Table 1: Notation table for DDIM sampling equation.