Summary: This presentation introduces h-Edit, a diffusion-based image editing framework that improves flexible and effective editing by formulating image editing as a reverse-time bridge modeling problem based on Doob’s h-transform. It first reviews image editing in Stable Diffusion through DDIM inversion and sampling, then discusses Edit Friendly residual and sampling steps before developing a diffusion bridge perspective. The method analyzes the role of the h-function theoretically and proposes an h-edit procedure using predictor-corrector sampling within Stable Diffusion. Experiments are presented to demonstrate the editing behavior, and the overall contribution is summarized as a reverse-time bridge framework for diffusion-based image editing with practical PC sampling.
Summary: This presentation introduces h-Edit, a diffusion-based image editing framework that improves flexible and effective editing by formulating image editing as a reverse-time bridge modeling problem based on Doob’s h-transform. It first reviews image editing in Stable Diffusion through DDIM inversion and sampling, then discusses Edit Friendly residual and sampling steps before developing a diffusion bridge perspective. The method analyzes the role of the h-function theoretically and proposes an h-edit procedure using predictor-corrector sampling within Stable Diffusion. Experiments are presented to demonstrate the editing behavior, and the overall contribution is summarized as a reverse-time bridge framework for diffusion-based image editing with practical PC sampling.