Research map / Image generation / Super-resolution
BUFF: Bayesian Uncertainty Guided Diffusion Probabilistic Model for Single Image Super-Resolution
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BUFF, which uses a Bayesian network's per-pixel uncertainty masks to modulate the Gaussian noise of an SRDiff-style diffusion model for single image super-resolution.
Problem
Diffusion-based super-resolution models assume i.i.d. Gaussian noise, which ignores varying texture complexity across image regions and leads to blur and artifacts at edges and fine details.
Approach
A Bayesian network predicts per-pixel mean and generalized-Gaussian scale/shape parameters; the derived uncertainty mask is nonlinearly refined (sigmoid-based amplification/reduction) and used to modulate the noise in the forward and reverse diffusion process of SRDiff.
Key results
- On BSD100 (4x), BUFF reports PSNR 26.40 and SSIM 0.7011, a +0.61 dB and +0.0198 SSIM gain over the SRDiff baseline.
- On DIV2K validation, BUFF reaches 29.35 PSNR vs 28.76 for SRDiff (+0.59 dB).
- Ablation on mask quality shows higher-quality (lower AUSE) uncertainty masks yield better Urban100 and DIV2K PSNR/SSIM/FID.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Bayesian uncertainty-guided noise modulation | SRDiff · Set14 | PSNR: +0.42 | Helped |
| add Bayesian uncertainty-guided noise modulation | SRDiff · Urban100 | PSNR: +0.36 | Helped |
| add Bayesian uncertainty-guided noise modulation | SRDiff · BSD100 | PSNR: +0.61 | Helped |
| add Bayesian uncertainty-guided noise modulation | SRDiff · Manga109 | PSNR: +0.56 | Helped |
| add Bayesian uncertainty-guided noise modulation | SRDiff · General100 | PSNR: +0.42 | Helped |
| add Bayesian uncertainty-guided noise modulation | SRDiff · DIV2K validation | PSNR: +0.59 | Helped |
| use a more accurate uncertainty mask (lower AUSE) (ablation) | BUFF · Urban100 | PSNR: 25.01 -> 25.49 | Helped |
| use a more accurate uncertainty mask (lower AUSE) (ablation) | BUFF · DIV2K | PSNR: 29.01 -> 29.35 | Helped |
| set amplification intensity to 1.30 instead of 1.20 (ablation) | BUFF · Urban100 | PSNR: 25.49 -> 25.41 | Hurt |
| set amplification intensity to 1.10 instead of 1.20 (ablation) | BUFF · DIV2K | PSNR: 29.35 -> 29.17 | Hurt |
| combine Bayesian Guided and Bayesian Embedding (ablation) | BUFF · BSD100 ×4 SR | PSNR: 26.21 -> 26.40 | Helped |
| add Bayesian Guided to Bayesian Embedding (ablation) | BUFF · BSD100 ×4 SR | PSNR: 26.01 -> 26.40 | Helped |
Surprising findings
- BUFF's SSIM is lower than some baselines on Set14 and BSD100 gains are mixed, with a marginal SSIM decrease noted on Set14 and BSD100 in the text, conflicting with the table's reported gains.
Limitations
- Paper acknowledges Bayesian training is computationally intensive and unstable, motivating the approach but adding cost.
- Ablation on amplification intensity shows sensitivity to hyperparameters (1.30 hurts, 1.10 gives little gain).
Experimental setup
Datasets: DIV2K Flickr2K FFHQ CelebA-HQ Set5 Set14 B100 Urban100 Manga109 General100 ImageNet-1K
Benchmarks: Set14 Urban100 BSD100 Manga109 General100 DIV2K DIV2K100 ImageNet-1K CelebA-HQ
Metrics: PSNR SSIM LPIPS FID AUSE
Baselines: SRCNN EDSR RCAN ESRGAN Real-ESRGAN BSRGAN BebyGAN SwinIR HAT IDM ACDMSR SAM-DiffSR
Builds on: SRDiff RRDB DDPM
Closest papers by meaning
- Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model (CVPR 2025 · similarity 0.92)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.90)
- Arbitrary-steps Image Super-resolution via Diffusion Inversion (CVPR 2025 · similarity 0.89)
- RAP-SR: RestorAtion Prior Enhancement in Diffusion Models for Realistic Image Super-Resolution (AAAI 2025 · similarity 0.89)
- BayesDiff: Estimating Pixel-wise Uncertainty in Diffusion via Bayesian Inference (ICLR 2024 · similarity 0.89)
- Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs (NeurIPS 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.