Research map / Image generation / Image restoration

DeblurDiff: Real-Word Image Deblurring with Generative Diffusion Models

NeurIPS 2025 · Image generation › Image restoration · approach: Diffusion priors for denoising and deblurring · 5 citations (OpenAlex)

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 DeblurDiff, which co-trains a latent kernel prediction network with a ControlNet-style conditional Stable Diffusion model, using element-wise adaptive convolution to produce clear latent guidance that is iteratively refined at each diffusion step for real-world image deblurring.

Problem

Real-world image deblurring with pre-trained Stable Diffusion priors is hindered when blurry images or pre-deblurred outputs from a separate deblurring network are used as conditions, since structure extraction is inaccurate and results depend heavily on that network.

Approach

A Latent Kernel Prediction Network, trained jointly with the diffusion model, predicts per-pixel kernels that are applied to the blurry latent via element-wise adaptive convolution (EAC) to yield clear latent guidance; the kernels are re-estimated from intermediate diffusion outputs at each sampling step.

Key results

• On the Real Blurry Images dataset, the ablation shows NIQE 3.6628 and MANIQA 0.5963 for full DeblurDiff versus 4.0978 and 0.5544 for the ControlNet baseline.
• On GoPro, DeblurDiff reports PSNR 24.32 and LPIPS 0.2191, below HI-Diff (PSNR 33.33, LPIPS 0.0820), while achieving the best MUSIQ (61.6369) and MANIQA (0.6134) among compared methods.
• On RWBI, DeblurDiff reaches MUSIQ 66.7505 and CLIP-IQA 0.6849, the best among compared methods; the paper notes HI-Diff is stronger on reference-based metrics.

Measured effects reported in the paper

Surprising findings

• The method scores lower than HI-Diff on reference-based metrics (PSNR, SSIM, LPIPS) on GoPro, DVD and RealBlur, while winning on no-reference metrics.

Limitations

• Lower reference-based fidelity (PSNR/SSIM/LPIPS) than regression-based HI-Diff on several benchmarks.
• Training dataset is private and large (about 500k pairs), limiting reproducibility of the training setup.
• Relies on retraining all competitors on the authors' own dataset, so comparisons depend on that setup.

Experimental setup

Datasets: MC-Blur RSBlur REDS-style self-captured video pairs synthetic motion blur on high-definition images GoPro DVD RealBlur RWBI Real Blurry Images

Benchmarks: GoPro DVD RealBlur RWBI Real Blurry Images

Metrics: PSNR SSIM LPIPS NIQE MUSIQ MANIQA CLIP-IQA

Baselines: FFTformer DBGAN ResShift HI-Diff ControlNet PASD DiffBIR

Builds on: Stable Diffusion ControlNet DDPM kernel prediction network FFTformer

Closest papers by meaning

• BlurDM: A Blur Diffusion Model for Image Deblurring (NeurIPS 2025 · similarity 0.90)
• Residual Diffusion Deblurring Model for Single Image Defocus Deblurring (AAAI 2025 · similarity 0.89)
• Prior and Prediction Inverse Kernel Transformer for Single Image Defocus Deblurring (AAAI 2024 · similarity 0.89)
• Exploiting Diffusion Prior for Real-World Image Dehazing with Unpaired Training (AAAI 2025 · similarity 0.89)
• Parameterized Blur Kernel Prior Learning for Local Motion Deblurring (CVPR 2025 · similarity 0.88)
• Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · similarity 0.88)