Research map / Image generation / Image restoration
Fourier Priors-Guided Diffusion for Zero-Shot Joint Low-Light Enhancement and Deblurring
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 FourierDiff, a zero-shot joint low-light enhancement and deblurring method that embeds Fourier amplitude/phase priors into a pre-trained ImageNet diffusion model without paired data or degradation assumptions.
Problem
Joint low-light enhancement and deblurring of real-world images; existing learned methods rely on paired synthetic data and generalize poorly, while diffusion zero-shot methods assume known degradation models.
Approach
Decompose sampled diffusion estimates in the Fourier domain: replace the phase with the input's phase for content consistency, and rescale/mix amplitudes so brightness follows the generative prior; additionally iteratively refine the input phase via blur-kernel estimation from the sampled results (spatial-frequency alternating optimization).
Key results
- On RealBlur, FourierDiff achieves best no-reference scores (e.g., NIQE 3.25, PI 3.36, BRISQUE 26.39, MUSIQ 52.24), outperforming all cascade baselines and LEDNet.
- On LOL-Blur, FourierDiff reaches PSNR 20.53 and SSIM 0.71 without paired training, comparable to LEDNet (25.74/0.85) which is trained on that dataset, and beats the other zero-shot baselines.
- In a 40-participant user study, the method was preferred over the compared baselines by large margins (e.g., 97.0% vs Zero-DCE++→GRL, 83.5% vs GDP→GRL).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Fourier priors-guided diffusion sampling and spatial-frequency alternating optimization | FourierDiff · LOL-Blur | MUSIQ: 59.64 -> 62.46 | Helped |
| add Fourier priors-guided diffusion sampling and spatial-frequency alternating optimization | FourierDiff · LOL-Blur | PSNR: 25.74 -> 20.53 | Hurt |
| add Fourier priors-guided diffusion sampling and spatial-frequency alternating optimization | FourierDiff · RealBlur night | BRISQUE: 28.96 -> 26.39 | Helped |
| add spatial-frequency alternating optimization (ablation) | FourierDiff · RealBlur night | NIQE: 3.62 -> 3.25 | Helped |
| add spatial-frequency alternating optimization (ablation) | FourierDiff · RealBlur night | MUSIQ: 48.81 -> 52.24 | Helped |
| set alternating optimization interval N to 500 instead of 200 (ablation) | FourierDiff · RealBlur night | NIQE: 3.25 -> 3.37 | Hurt |
| set alternating optimization interval N to 500 instead of 200 (ablation) | FourierDiff · RealBlur night | MUSIQ: 52.24 -> 49.87 | Hurt |
| set alternating optimization interval N to 100 instead of 200 (ablation) | FourierDiff · RealBlur night | NIQE: 3.25 -> 3.26 | No clear effect |
| set alternating optimization interval N to 100 instead of 200 (ablation) | FourierDiff · RealBlur night | MUSIQ: 52.24 -> 52.41 | No clear effect |
| set alternating optimization interval N to 1 instead of 200 (ablation) | FourierDiff · RealBlur night | NIQE: 3.25 -> 3.19 | Helped |
| set alternating optimization interval N to 1 instead of 200 (ablation) | FourierDiff · RealBlur night | MUSIQ: 52.24 -> 51.11 | Hurt |
Surprising findings
- Pure LEDNet trained on LOL-Blur attains higher PSNR/SSIM (25.74/0.85) than FourierDiff on LOL-Blur, which the authors acknowledge only as 'comparable'.
- Brightness adjustment is not strictly needed: the ImageNet diffusion prior alone yields reasonable brightness (w/o BA).
Limitations
- Performs poorly in extremely dark environments due to loss of content guidance.
- Inherits slow diffusion sampling; not real-time.
- Relies on blur kernel estimation which is inaccurate in low light.
Experimental setup
Datasets: LOL-Blur RealBlur ImageNet
Benchmarks: LOL-Blur RealBlur
Metrics: NIQE PI BRISQUE MUSIQ PSNR SSIM
Baselines: LEDNet Zero-DCE++ RetinexDIP GDP GRL Chen et al. low-light deblurring W-DIP
Builds on: DDRM DDNM GDP Denoising Diffusion Probabilistic Models
Closest papers by meaning
- AGLLDiff: Guiding Diffusion Models Towards Unsupervised Training-free Real-world Low-light Image Enhancement (AAAI 2025 · similarity 0.90)
- Zero-Reference Low-Light Enhancement via Physical Quadruple Priors (CVPR 2024 · similarity 0.88)
- Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models (AAAI 2025 · similarity 0.88)
- ZERO-IG: Zero-Shot Illumination-Guided Joint Denoising and Adaptive Enhancement for Low-Light Images (CVPR 2024 · similarity 0.88)
- Multi-Domain Multi-Scale Diffusion Model for Low-Light Image Enhancement (AAAI 2024 · similarity 0.88)
- Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios (ICLR 2025 · similarity 0.87)
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.