Diffusion-based Deblurring
Removes motion blur from images using generative diffusion models with Stable Diffusion priors. Uses a Latent Kernel Prediction Network (LKPN) for robust real-world deblurring with iterative refinement. Achieves state-of-the-art results on both benchmark and real-world images with superior detail reconstruction.
What it does
DeblurDiff is a deep learning model that removes motion blur from photographs using generative diffusion models. It introduces a Latent Kernel Prediction Network (LKPN) that learns spatially variant kernels to guide sharp image restoration in latent space. The model co-trains with Stable Diffusion to leverage clear image priors while using Element-wise Adaptive Convolution (EAC) to preserve structural information from the blurry input.
Problem it solves
- Remove motion blur from real-world photographs
- Restore sharpness while preserving structural integrity
- Handle complex, spatially variant blur patterns
- Reconstruct realistic textures and fine details
- Overcome limitations of pre-deblurred conditioning approaches
- Generalize across different blur types and domains
Input/Output
Accuracy & Speed
- Performance:
- Accuracy: Outperforms state-of-the-art methods on benchmark and real-world images
- Training datasets: Synthetic blur-sharp pairs with real-world validation
- Comparisons: Exceeds FFTformer, DBGAN, HiDiff, ResShift, ControlNet, PASD, and DiffBIR
- Metrics:
- Superior structural fidelity vs. competing methods
- High-quality detail and texture reconstruction
- Robust performance on unseen blur patterns
Technical Details
- Architecture: Latent Kernel Prediction Network (LKPN) + Stable Diffusion
- Core Innovation:
- LKPN learns spatially variant kernels in latent space
- Element-wise Adaptive Convolution (EAC) preserves structural information
- Iterative kernel refinement using intermediate diffusion results
- Co-training with conditional diffusion for robust guidance
- Design Benefits:
- Avoids dependency on pre-deblurred image quality
- Better generalization across blur types than DRM-based methods
- Preserves input structure while generating realistic details
- Iterative refinement improves accuracy progressively
- Framework & Dependencies:
- Framework: PyTorch
- Base model: Stable Diffusion (pre-trained)
- Dependencies: See environment.yml in repository
- GPU required for practical inference times
- Training:
- Approach: Co-training LKPN with conditional diffusion
- Learning strategy: Iterative kernel estimation with diffusion feedback
- Built upon: DiffBIR and DemystifyLocalViT codebases
Compliance & Provenance
Provider | Open-source |
Provider type | Specialized |
License | MIT |
EU AI Act risk class | Minimal Risk |
Art. 50 transparency | Not applicable |
Region availability | Available globally |
Training data summary | Pending — provider has not yet published per Art. 53(d) |
For more on how we classify models and mark outputs, see our AI Policy.
Model Source
- GitHub Repository: https://github.com/kkkls/DeblurDiff
- Paper: "DeblurDiff: Real-World Image Deblurring with Generative Diffusion Models" NeurIPS 2025" https://arxiv.org/abs/2502.03810
- License: MIT License