Stable and Near-Reversible Diffusion ODE Solvers for Image Editing

TL;DR

This paper introduces near-reversible diffusion ODE solvers with Runge-Kutta methods and vector-field smoothing, enhancing stability and edit quality in image editing compared to purely reversible approaches.

Contribution

It proposes a near-reversible approach combining Runge-Kutta methods and smoothing to improve stability and fidelity in diffusion-based image editing.

Findings

Near-reversible solvers improve edit fidelity.

The approach remains stable under large edits.

Background preservation benefits are retained.

Abstract

The inversion of diffusion models plays a central role in image editing. Algebraically reversible ODE solvers provide an appealing approach to diffusion inversion for text-guided image editing, by eliminating the inversion error inherent in DDIM-based editing pipelines. However, empirical results indicate that reversibility alone is insufficient. As edits require larger semantic or visual changes, reversible diffusion solvers often exhibit instabilities and suffer sharp drops in output quality. In this paper, we show that the trade-off between exact reversibility and numerical stability manifests empirically as a trade-off between background preservation and prompt alignment in image editing. We then investigate the use of near-reversible Runge-Kutta methods as a more stable alternative to exactly reversible diffusion schemes. When combined with a vector-field smoothing strategy, the resulting approach improves edit fidelity, remains stable under large edits, and largely retains the background-preservation benefits of reversible solvers.