The article presents a unified probabilistic formulation for diffusion-based image editing. It demonstrates that Stochastic Differential Equations (SDEs) offer significant improvements over Ordinary Differential Equations (ODEs) in various image editing tasks. The study introduces SDE-Drag, a new method for point-based content dragging, and a challenging benchmark, DragBench, for evaluation. User studies indicate that SDE-Drag outperforms the ODE baseline, existing diffusion-based methods, and the renowned DragGAN. The results underline the versatility and superiority of SDE in image editing.
Publication date: 2 Nov 2023
Project Page: https://ml-gsai.github.io/SDE-Drag-demo/
Paper: https://arxiv.org/pdf/2311.01410
