MeanSR: Restoration Trajectory Learning for One-Step Perceptual Super-Resolution

2026-08-10

AI Summary

This work addresses the limitations of existing single-step super-resolution methods, which either rely on computationally expensive pretrained teacher models or fail to explicitly model the dynamic process of image restoration from low- to high-resolution. The paper proposes MeanSR, a novel framework that, for the first time, directly models the finite-time image restoration trajectory without knowledge distillation by learning a mean velocity field conditioned on low-resolution inputs. MeanSR integrates ODE-based trajectory generation, distributional trajectory matching, and a stage-aware temporal sampling strategy to achieve a favorable trade-off between efficiency and perceptual quality. Experiments demonstrate that MeanSR outperforms CTMSR on perceptual metrics including CLIPIQA, MUSIQ, and MANIQA, while significantly reducing FLOPs and inference latency. The reconstructed images exhibit sharper structures, more realistic textures, and fewer perceptual artifacts.

Abstract

Diffusion-based super-resolution (SR) achieves strong perceptual quality but requires costly iterative denoising. Existing one-step distillation methods reduce inference time but depend on expensive pretrained teachers, whereas CTMSR avoids distillation through PF-ODE consistency training yet does not explicitly model the restoration dynamics from low-resolution (LR) inputs to high-resolution (HR) images. We propose MeanSR, a one-step perceptual SR method that learns an LR-conditioned average velocity field to directly capture the finite-time transition from degraded or noisy inputs to plausible HR outputs. We further reformulate distribution trajectory matching for average-velocity generation and introduce a Stage-Aware Temporal Sampling strategy to improve trajectory learning. Experiments on synthetic and real-world benchmarks show that MeanSR outperforms CTMSR on CLIPIQA, MUSIQ, and MANIQA while substantially reducing FLOPs and inference latency. MeanSR also reconstructs sharper structures and more realistic textures with fewer perceptual artifacts.

Problem

• super-resolution
• perceptual quality
• one-step generation
• restoration dynamics
• diffusion models

Innovation

• one-step super-resolution
• average velocity field
• trajectory learning
• perceptual quality
• diffusion model