Title: Accelerating Image Super-Resolution Through Consistency Rectified Flow

Abstract:

Although diffusion models (DMs) have achieved significant breakthroughs in practical image super-resolution (SR), their deployment is often impeded by the computational burden of multi-step sampling. While recent studies have proposed few- or single-step alternatives, these approaches frequently suffer from inefficient modeling of noisy inputs or an inability to fully leverage iterative generative priors, which ultimately degrades the fidelity and quality of the output. To overcome these limitations, we introduce FlowSR, a new framework that redefines SR as a rectified flow transitioning from low-resolution (LR) to high-resolution (HR) images. This method employs an enhanced consistency learning strategy to facilitate high-quality SR in just one step. We improve upon standard consistency distillation by integrating HR regularization, a technique that guarantees the learned SR flow maintains self-consistency while accurately converging to the ground-truth HR target. Additionally, we propose a fast-slow scheduling mechanism for consistency learning, where adjacent timesteps are drawn from two separate schedulers: a fast scheduler with reduced timesteps to boost efficiency, and a slow scheduler with increased timesteps to preserve fine-grained texture details. Comprehensive experiments confirm that FlowSR delivers superior results in both computational efficiency and image quality.

Code: \href{https://github.com/jiaqixuac/FlowSR}{this https URL}.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC