POD-SR: Pixel-Space One-Step Diffusion with Degradation-Aware Timestep and Trajectory Alignment for Super-Resolution
Abstract
Diffusion models have demonstrated superior performance in image super-resolution, yet balancing perceptual quality, pixel-level fidelity, and inference speed remains a significant challenge. Most existing approaches prioritize efficiency by leveraging Variational Autoencoders (VAEs) for latent-space generation; however, this paradigm introduces an inherent information bottleneck that limits pixel-level fidelity, particularly when recovering strongly semantic details such as small text. To address this trade-off, we propose POD-SR, a high-fidelity, pixel-space one-step diffusion framework. POD-SR utilizes a frequency-decoupled pixel-space architecture to circumvent the VAE bottleneck, effectively preserving high-frequency details. Furthermore, to maintain restorative quality in a single step, we introduce two key modules: a Degradation-Aware Graph Perception Network (D-GPN) and an Asymmetric Generative Prior Alignment (AGPA) module. Specifically, D-GPN models local structural anomalies to dynamically predict image-specific starting timesteps, providing an adaptive alternative to static timesteps. Concurrently, AGPA realigns the restorative trajectory with the model's intrinsic generative prior to enhance structural faithfulness. Extensive evaluations confirm that POD-SR achieves competitive performance in fidelity metrics while maintaining competitive inference efficiency.