Denoising Diffusion Probabilistic Models for Robust Image Super-Resolution in the Wild

TL;DR

SR3+ addresses blind single-image super-resolution in the wild, where degradations are unknown and out-of-distribution. It combines a diffusion-based denoiser with two innovations: higher-order degradations during training and noise conditioning augmentation, enabling robust, texture-rich reconstructions. The approach achieves state-of-the-art FID on RealSR and DRealSR zero-shot benchmarks, outperforming Real-ESRGAN and SR3 when trained on comparable data, with further gains from larger models and datasets. It also provides a tunable test-time mechanism (t_eval) to balance fidelity to the input and perceptual realism, paving the way for more robust diffusion-based super-resolution in practical, uncontrolled settings.

Abstract

Diffusion models have shown promising results on single-image super-resolution and other image- to-image translation tasks. Despite this success, they have not outperformed state-of-the-art GAN models on the more challenging blind super-resolution task, where the input images are out of distribution, with unknown degradations. This paper introduces SR3+, a diffusion-based model for blind super-resolution, establishing a new state-of-the-art. To this end, we advocate self-supervised training with a combination of composite, parameterized degradations for self-supervised training, and noise-conditioing augmentation during training and testing. With these innovations, a large-scale convolutional architecture, and large-scale datasets, SR3+ greatly outperforms SR3. It outperforms Real-ESRGAN when trained on the same data, with a DRealSR FID score of 36.82 vs. 37.22, which further improves to FID of 32.37 with larger models, and further still with larger training sets.

Figures (11)

• Figure 1: Blind super-resolution test results ($64\!\times\!64 \rightarrow 256\!\times\! 256$) for SR3+, SR3 and Real-ESRGAN.
• Figure 2: The SR3+ data pipeline applies a sequence of degradations to HR training images (like Real-ESRGAN but without additive noise). To form the conditioning signal for the neural denoiser, we up-sample the LR image and applied noise conditioning augmentation.
• Figure 3: Sample comparison between Real-ESRGAN and various SR3+ models (ours). We observe that Real-ESRGAN often suffers from oversmoothing and excessive contrast, while SR3+ is capable of generating high-fidelity, realistic textures.
• Figure 4: Ablation samples ($t_{eval}\!=\! 0.1$), illustrating the importance of higher-order degradations and noise conditioning augmentation.
• Figure 5: Samples from SR3+ (400M weights, 61M dataset) using different amounts of test-time noise conditioning augmentation, $t_{eval}$.