An Efficient Uncertainty-Driven Learning for Stochastic Super-Resolution

초록

Single image super-resolution (SISR) is inherently ill-posed because of information loss during degradation, where a single low-resolution (LR) image corresponds to multiple plausible high-resolution (HR) reconstructions. While recent stochastic SISR approaches address this by modeling conditional distributions, they commonly assume isotropic Gaussian priors, overlooking spatial variations in uncertainty across the image. In this paper, we propose a novel uncertainty-driven framework for stochastic SISR that incorporates anisotropic Gaussian priors modulated by estimated uncertainty maps. Our model explicitly learns the spatially varying aleatoric uncertainty from LR inputs and uses it to guide the sampling of latent noise, which is then injected into a decoder to reconstruct residuals over a deterministic SISR baseline. This approach not only captures the one-to-many nature of the SISR task but also preserves strong pixel-level fidelity by centering the generation around a pre-trained deterministic prediction. Extensive experiments show that our method achieves state-of-the-art performance in terms of both distortion and perceptual quality, especially outperforming the best generative-model-based methods by +0.44 in PSNR and -0.008 in LPIPS simultaneously with lower computational cost.

키워드