Research map / Image generation / Super-resolution
A Dynamic Kernel Prior Model for Unsupervised Blind 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 DKP, an unsupervised, pre-training-free kernel estimation module (MCMC-sampled kernel priors plus network-based Langevin optimization) that plugs into DIP and diffusion restorers for blind SR.
Problem
Blind super-resolution requires estimating unknown blur kernels, but leading learning-based methods need supervised pre-training on labeled LR-HR or kernel datasets, limiting generalization across kernel types.
Approach
A Dynamic Kernel Prior: kernel priors are obtained by likelihood-weighted Monte Carlo (MCMC-style) sampling of random kernels given the current HR estimate and LR input, and a lightweight kernel network is updated via a network-based Langevin dynamics step that uses these priors as noise-like perturbations, all without pre-training.
Key results
- Diff-DKP achieves 29.44 dB PSNR on Set5 (x4, Gaussian kernels), the best among compared methods, versus 27.81 for BSRDM.
- On motion kernels, Diff-DKP reaches 28.74 dB on Set5 versus 24.78 for DCLS, showing stable performance across kernel types.
- DIP-DKP and Diff-DKP reach 56.20 and 56.76 kernel PSNR on Set14 Gaussian, above BSRDM's 55.38; DIP-DKP runtime is about 92 s and 11 GB for a 512x512 image, comparable to Double-DIP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Diff-DKP instead of the strongest reported baseline | Blind super-resolution · Gaussian kernels, Set5/Set14/BSD100/Urban100 x4 | PSNR: 27.81 -> 29.44; 25.68 -> 26.76; 25.61 -> 26.63; 22.92 -> 23.92 | Helped |
| Use DIP-DKP instead of the strongest reported baseline | Blind super-resolution · Gaussian kernels, Set5/Set14/BSD100/Urban100 x4 | PSNR: 27.81 -> 28.03; 25.68 -> 25.98; 25.61 -> 25.66; 22.92 -> 23.24 | Helped |
| Use Diff-DKP instead of the strongest reported baseline | Blind super-resolution · Motion kernels, Set5/Set14/BSD100/Urban100 x4 | PSNR: 24.78 -> 28.74; 24.38 -> 26.03; 22.80 -> 24.10; 20.58 -> 22.26 | Helped |
| Use DIP-DKP instead of the strongest reported baseline | Blind super-resolution · Motion kernels, Set5/Set14/BSD100/Urban100 x4 | PSNR: 24.78 -> 25.30; 24.38 -> 24.52; 22.80 -> 23.02; 20.58 -> 21.24 | Helped |
| Remove the RKS module (ablation) | DIP-DKP · Gaussian kernels, Set14 x4 | image PSNR: 25.98 -> 18.77 | Hurt |
| Remove the RKS module (ablation) | Diff-DKP · Gaussian kernels, Set14 x4 | image PSNR: 26.76 -> 17.33 | Hurt |
| Remove the RKS module (ablation) | DIP-DKP · Motion kernels, Set14 x4 | image PSNR: 24.52 -> 18.19 | Hurt |
| Remove the RKS module (ablation) | Diff-DKP · Motion kernels, Set14 x4 | image PSNR: 26.03 -> 17.65 | Hurt |
| Vary MCMC sampling count L (ablation) | DIP-DKP · Set5, Gaussian kernel | image PSNR: 20.99; 27.12; 28.44; 28.57; 28.52; 28.29; 28.03 | Mixed |
| Vary MCMC sampling count L (ablation) | Diff-DKP · Set5, Gaussian kernel | image PSNR: 21.97; 28.95; 29.40; 29.47; 29.76; 29.67; 29.26 | Mixed |
| Vary PKE kernel-network depth and units (ablation) | DIP-DKP · Set5 x4 | image PSNR: 13.75; 23.57; 28.93; 28.24; 13.61; 28.97; 28.48; 28.35; 13.30; 28.81; 28.52; 26.65; 13.86; 28.30; 28.54; 27.93 | Mixed |
Surprising findings
- Without the RKS module, Diff-DKP's motion kernel estimation fails to converge, with Set14 image PSNR dropping to 17.65 dB.
- Pre-trained kernel networks (FKP) do not generalize well when kernel categories vary, per the authors.
Limitations
- Inference is per-image optimization, so runtime (about 60-92 s per 512x512 image) is non-trivial.
- DIP-DKP gains over existing methods are only slight on Gaussian kernels.
- Evaluation is restricted to x4 scale and synthetic Gaussian/motion kernels, with only qualitative real-image results.
Experimental setup
Datasets: Set5 Set14 BSD100 Urban100 RealSRSet
Benchmarks: Set5 Set14 BSD100 Urban100 RealSRSet
Metrics: PSNR SSIM kernel PSNR
Baselines: Double-DIP DIP-FKP DASR BSRDM DCLS DiffBIR DARM
Builds on: Deep Image Prior Double-DIP FKP DDPM
Closest papers by meaning
- Flow-Based Kernel Prior With Application to Blind Super-Resolution (CVPR 2021 · similarity 0.91)
- Deep Constrained Least Squares for Blind Image Super-Resolution (CVPR 2022 · similarity 0.91)
- ResDiff: Combining CNN and Diffusion Model for Image Super-resolution (AAAI 2024 · similarity 0.86)
- Noise-Free Optimization in Early Training Steps for Image Super-resolution (AAAI 2024 · similarity 0.85)
- A Modular Conditional Diffusion Framework for Image Reconstruction (NeurIPS 2024 · similarity 0.85)
- Taming Generative Diffusion Prior for Universal Blind Image Restoration (NeurIPS 2024 · similarity 0.85)
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.