Research map / Image generation / Super-resolution
ResDiff: Combining CNN and Diffusion Model for 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 ResDiff, a residual diffusion model for single image super-resolution where a pre-trained CNN (SimpleSR) gives an initial prediction and a frequency-domain-guided diffusion model predicts the residual to the ground truth.
Problem
Diffusion-based single image super-resolution must recover both low- and high-frequency content from noise, which slows convergence and wastes capacity on low-frequency content that a simple CNN can recover.
Approach
Use a pre-trained CNN's prediction as the starting point so the diffusion process targets the residual between ground truth and CNN output; the CNN is trained with added FFT and DWT frequency losses, and the diffusion U-Net uses an FD Info Splitter and high-frequency guided cross-attention built from DWT coefficients of the CNN output.
Key results
- On FFHQ 32->128 (4x), ResDiff reaches 26.73 PSNR, 0.818 SSIM and 70.54 FID, beating SRDiff (26.07/0.794/72.36) and SR3 (25.37/0.778/75.29).
- On DIV2K 4x, ResDiff reports 27.94 PSNR and 106.71 FID versus SRDiff's 26.87 and 110.32.
- On CelebA 64->256, ResDiff reaches 27.16 PSNR and 38.47 FID, better than SRDiff's 26.84 and 39.16.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ResDiff instead of SRDiff | FFHQ 32→128 (4×) · FFHQ 32→128 (4×) | PSNR: 26.07 vs 26.73 | Helped |
| use ResDiff instead of SRDiff | FFHQ 256→1024 (4×) · FFHQ 256→1024 (4×) | PSNR: 23.01 vs 23.15 | Helped |
| use ResDiff instead of SRDiff | CelebA 20→160 (8×) · CelebA 20→160 (8×) | PSNR: 25.32 vs 25.37 | Helped |
| use ResDiff instead of SRDiff | CelebA 64→256 (4×) · CelebA 64→256 (4×) | PSNR: 26.84 vs 27.16 | Helped |
| add the residual structure rather than use no residual structure (ablation) | ResDiff · FFHQ 4× | PSNR: 25.49 -> 26.73 | Helped |
| use SimpleSR instead of bilinear interpolation for initial prediction (ablation) | ResDiff · FFHQ 4× | PSNR: 25.99 -> 26.73 | Helped |
| use SimpleSR instead of SRCNN for initial prediction (ablation) | ResDiff · FFHQ 4× | PSNR: 26.14 -> 26.73 | Helped |
| add the FFT and DWT losses to GT-only CNN pretraining (ablation) | ResDiff · FFHQ 4× | PSNR: 26.47 -> 26.73 | Helped |
| add FD Info Splitter and HF-guided CA (ablation) | ResDiff · FFHQ 4× | PSNR: 25.41 -> 26.73 | Helped |
| add FD Info Splitter (ablation) | ResDiff with HF-guided CA · FFHQ 4× | PSNR: 26.09 -> 26.73 | Helped |
| add HF-guided CA (ablation) | ResDiff with FD Info Splitter · FFHQ 4× | PSNR: 25.97 -> 26.73 | Helped |
Surprising findings
- The paper reports that ResDiff does not outperform current SOTA SISR methods, attributing this to model size differences.
Limitations
- DWT operations are time-consuming
- Color discrepancy when the model is under-trained
- Does not outperform current SOTA SISR methods due to smaller U-Net
Experimental setup
Datasets: FFHQ CelebA DIV2K Urban100
Benchmarks: FFHQ CelebA DIV2K Urban100
Metrics: PSNR SSIM FID
Baselines: SRGAN ESRGAN BRGM PULSE SRDiff SR3 SRFlow SRCNN
Builds on: SR3 SRDiff DDPM Imagen SRCNN
Closest papers by meaning
- TSD-SR: One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution (CVPR 2025 · similarity 0.91)
- One-Step Effective Diffusion Network for Real-World Image Super-Resolution (NeurIPS 2024 · similarity 0.91)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.90)
- Effective Diffusion Transformer Architecture for Image Super-Resolution (AAAI 2025 · similarity 0.90)
- FaithDiff: Unleashing Diffusion Priors for Faithful Image Super-resolution (CVPR 2025 · similarity 0.89)
- Arbitrary-steps Image Super-resolution via Diffusion Inversion (CVPR 2025 · 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.