DriftSR: One-Step Real-World Image Super-Resolution via Distribution Drifting

One-step real-world image super-resolution (Real-ISR) offers efficient inference, but recovering realistic and perceptually rich details often relies on score distillation or adversarial learning, introducing additional trainable components and making optimization more cumbersome. To this end, we propose DriftSR, a one-step Real-ISR framework that leverages pretrained diffusion priors through distribution drifting. We construct the drifting space directly from frozen intermediate representations of a pretrained diffusion model, eliminating the need for task-specific feature encoders. We introduce Spatial Feature Drifting, which extends distribution modeling from image-level representations to spatial feature samples, providing denser supervision for local texture recovery. To prevent drifting from generating perceptually plausible but structurally inconsistent details, we propose Structure-Modulated Guidance, which adaptively modulates drifting supervision based on local structural consistency with the low-quality input. With these designs, DriftSR optimizes only the one-step generator, without auxiliary distillation branches or discriminators, while exploiting diffusion priors for perceptual detail enhancement. Extensive experiments on three real-world benchmarks demonstrate that DriftSR achieves strong perceptual quality