BUFF: Bayesian Uncertainty Guided Diffusion Probabilistic Model for Single Image Super-Resolution

TL;DR

BUFF tackles the problem of single-image SR by integrating Bayesian uncertainty into diffusion models to guide region-specific noise during reconstruction.A Bayesian neural network predicts per-pixel mean and uncertainty, yielding a mask $M_{Bayes}$ that is refined to $B$ and used to modulate the diffusion noise via forward process $oldsymbol{x}_t=\sqrt{1-\beta_t}\boldsymbol{x}_{t-1}+\sqrt{\beta_t}(\boldsymbol{\epsilon}_t\odot B)$ and the reverse posterior with $\tilde{\mu}$, enabling adaptive, context-aware restoration.Empirical results on DF2K/diverse benchmarks show BUFF outperforms diffusion baselines in PSNR/SSIM and demonstrates robustness to blur and facial detail restoration, with ablations highlighting the importance of uncertainty mask quality and noise modulation intensity.The approach represents a principled blend of Bayesian uncertainty estimation and diffusion-based SR, offering improved fidelity, fewer artifacts, and broader applicability to related restoration tasks.

Abstract

Super-resolution (SR) techniques are critical for enhancing image quality, particularly in scenarios where high-resolution imagery is essential yet limited by hardware constraints. Existing diffusion models for SR have relied predominantly on Gaussian models for noise generation, which often fall short when dealing with the complex and variable texture inherent in natural scenes. To address these deficiencies, we introduce the Bayesian Uncertainty Guided Diffusion Probabilistic Model (BUFF). BUFF distinguishes itself by incorporating a Bayesian network to generate high-resolution uncertainty masks. These masks guide the diffusion process, allowing for the adjustment of noise intensity in a manner that is both context-aware and adaptive. This novel approach not only enhances the fidelity of super-resolved images to their original high-resolution counterparts but also significantly mitigates artifacts and blurring in areas characterized by complex textures and fine details. The model demonstrates exceptional robustness against complex noise patterns and showcases superior adaptability in handling textures and edges within images. Empirical evidence, supported by visual results, illustrates the model's robustness, especially in challenging scenarios, and its effectiveness in addressing common SR issues such as blurring. Experimental evaluations conducted on the DIV2K dataset reveal that BUFF achieves a notable improvement, with a +0.61 increase compared to baseline in SSIM on BSD100, surpassing traditional diffusion approaches by an average additional +0.20dB PSNR gain. These findings underscore the potential of Bayesian methods in enhancing diffusion processes for SR, paving the way for future advancements in the field.

Figures (5)

• Figure 1: Comparison of the noise addition strategies in the forward diffusion phase between SRDiffsrdiff and BUFF approaches, with BUFF utilizing a Bayesian model to generate uncertainty masks that guide noise intensity adjustments across different regions of the image for SR results.
• Figure 2: Process diagram showcasing the Bayesian modeling for uncertainty estimation in image super-resolution, where a Bayesian neural network refines uncertainty measures to guide noise modulation, enhancing the LR to HR reconstruction process.
• Figure 3: Visualization of restored images generated by different methods. Our BUFF surpasses other approaches in terms of both higher reconstruction quality and fewer artifacts. Additional visualization results can be found in our supplementary material.
• Figure 4: Visual comparisons on CelebA-HQ dataset for 8× face SR. Zoom in for a better view.
• Figure 5: Visualization of the effect of uncertainty Mask Quality.