Research map / Image generation / Super-resolution
CDFormer: When Degradation Prediction Embraces Diffusion Model for 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 CDFormer, a blind super-resolution method that learns a Content Degradation Prior (CDP) with a diffusion-based estimator operating on a 1-D latent vector and injects it into a transformer SR network.
Problem
Blind image super-resolution with complex real-world degradations: prior methods estimate only kernel or degradation information and ignore content details, while diffusion-based SR is slow and produces texture errors.
Approach
Use a diffusion model as a low-cost estimator (4 sampling steps, ~3M params) that recreates a one-dimensional Content Degradation Prior from LR images alone, trained in two stages with a GT encoder on HR/LR pairs, and inject the prior into a transformer SR network via learnable affine modulation (CDIM) and an interflow of spatial/channel attention and depth-wise convolution branches.
Key results
- On Urban100 at x4 with kernel width 0, CDFormer reaches 27.21 dB PSNR versus 26.50 for DCLS, and up to 0.71 dB gains over DCLS in blind settings.
- Compared with StableSR and SR3 on Set14 x4, CDFormer gets 29.00 dB PSNR (CDFormer-S 28.80) with 24.46M params and 0.39 s, versus StableSR's 19.66 dB at 1409M params and 71 s.
- Ablation on Set5 shows the combined degradation and content prior (model4) gives 32.564 dB vs 32.074 dB with neither prior at kernel width 1.2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add degradation prior to the no-prior model (ablation) | CDFormer · Set5, ×4, kernel width 1.2 | PSNR: 32.074 -> 32.372 | Helped |
| Add content prior to the no-prior model (ablation) | CDFormer · Set5, ×4, kernel width 1.2 | PSNR: 32.074 -> 32.451 | Helped |
| Use both degradation and content priors instead of no prior (ablation) | CDFormer · Set5, ×4, kernel width 1.2 | PSNR: 32.074 -> 32.564 | Helped |
| Add degradation prior to the no-prior model (ablation) | CDFormer · Set5, ×4, kernel width 2.4 | PSNR: 32.006 -> 32.167 | Helped |
| Add content prior to the no-prior model (ablation) | CDFormer · Set5, ×4, kernel width 2.4 | PSNR: 32.006 -> 32.287 | Helped |
| Use both degradation and content priors instead of no prior (ablation) | CDFormer · Set5, ×4, kernel width 2.4 | PSNR: 32.006 -> 32.393 | Helped |
| Generate CDP with the Stage 2 diffusion model rather than use Stage 1 (ablation) | CDFormer · Urban100, ×4, kernel width 0 | PSNR: 26.945 -> 27.210 | Helped |
| Generate CDP with the Stage 2 diffusion model rather than use Stage 1 (ablation) | CDFormer · Urban100, ×4, kernel width 1.2 | PSNR: 26.923 -> 27.058 | Helped |
| Use CDFormer instead of DCLS | Blind image super-resolution · Urban100, ×4, kernel width 1.2, noise-free | PSNR: 26.50 -> 27.21 | Helped |
Surprising findings
- Diffusion-generated CDP from LR images alone (stage 2) slightly outperforms the ground-truth-encoder version (stage 1) on Urban100, e.g. 27.210 vs 26.945 dB at kernel width 0.
Limitations
- Gains shrink as noise level increases, since severely degraded LR images carry too little information.
- Some large prior models could not be run on the single RTX 4090 in memory-heavy settings, leaving blank comparisons.
- Comparisons with StableSR and SR3 are limited to noise-free isotropic kernels in the main table.
Experimental setup
Datasets: DIV2K Flickr2K
Benchmarks: Set5 Set14 B100 Urban100
Metrics: PSNR SSIM FID LPIPS
Baselines: SR3 StableSR DCLS DASR KDSR RCAN
Builds on: DASR KDSR DCLS DiffIR DDPM
Closest papers by meaning
- Unsupervised Degradation Representation Aware Transform for Real-World Blind Image Super-Resolution (AAAI 2025 · similarity 0.89)
- Unsupervised Diffusion-Based Degradation Modeling for Real-World Super-Resolution (AAAI 2025 · similarity 0.88)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.86)
- Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs (NeurIPS 2024 · similarity 0.86)
- Effective Diffusion Transformer Architecture for Image Super-Resolution (AAAI 2025 · similarity 0.86)
- Unleashing the Power of One-Step Diffusion based Image Super-Resolution via a Large-Scale Diffusion Discriminator (NeurIPS 2025 · similarity 0.86)
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.