cs.CVMar 15, 2026

BluRef: Unsupervised Image Deblurring with Dense-Matching References

Authors: Bang-Dang Pham, Anh Tran, Cuong Pham, Minh Hoai

Organizations: University of Wisconsin-Madison · Qualcomm AI Research · Posts & Telecommunications Inst. of Tech. · Adelaide University

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

This paper introduces a novel unsupervised approach for image deblurring that utilizes a simple process for training data collection, thereby enhancing the applicability and effectiveness of deblurring methods. Our technique does not require meticulously paired data of blurred and corresponding sharp images; instead, it uses unpaired blurred and sharp images of similar scenes to generate pseudo-ground truth data by leveraging a dense matching model to identify correspondences between a blurry image and reference sharp images. Thanks to the simplicity of the training data collection process, our approach does not rely on existing paired training data or pre-trained networks, making it more adaptable to various scenarios and suitable for networks of different sizes, including those designed for low-resource devices. We demonstrate that this novel approach achieves state-of-the-art performance, marking a significant advancement in the field of image deblurring.

Figures & tables

Figure 1: Comparison of three approaches to image deblurring. Supervised deblurring requires costly paired data, limiting scalability. Reblurring-based methods rely on indirect mappings and suitable intermediate domains, which can be hard to identify. In contrast, the proposed BluRef directly learns from unpaired blurry and reference images within the target domain, offering an efficient and scalable solution without requiring paired supervision or complex multi-stage pipelines.
Figure 2: Iterative approach for generating a pseudo-sharp image and using it to train the deblurring model. In each epoch, the reference images {Irefn}n=1N\{I_{\text{ref}}^{n}\}_{n=1}^{N} are matched with the deblurred result Ideblur(k)I_{\text{deblur}}^{(k)} to produce the updated pseudo-sharp image and the corresponding confidence mask, which serve as the supervision targets for training the deblurring model.
Figure 3: Protocol to collect sharp reference images in our experiments. For each blurry image, we collect NN sharp reference images as the two groups of consecutive frames from both sides, each displaced by Δ\Delta frames from the blurry image.
 GoProRB2V
Delta ( Δ\Delta )1 frame10 frames20 frames1 frame10 frames20 frames
Unsupervised Deblurring      
DualGAN22.23 / 0.72122.10 / 0.71921.24 / 0.70221.01 / 0.51220.87 / 0.50020.92 / 0.505
UID-GAN23.42 / 0.73223.18 / 0.72422.38 / 0.72422.22 / 0.57822.01 / 0.55122.13 / 0.569
UAUD24.25 / 0.79224.02 / 0.75023.77 / 0.74522.87 / 0.59022.29 / 0.58122.28 / 0.581
BluRef (Ours)      
Table 1: Comparison of different deblurring methods on GoPro and RB2V datasets. For each test, we report PSNR ↑\uparrow / SSIM ↑\uparrow scores as evaluation metrics. The best scores are in bold and the second best score are in underline . For a supervised method, NAFNet or Restormer, we assess its upperbound of deblurring performance by training it on the paired dataset.
Figure 4: Qualitative results of several methods on GoPro and RB2V data. This figure presents four examples, two from each dataset. For each example, a zoomed-in portion of the input image is displayed, accompanied by its ground truth counterpart and the results from BluRef and three other baselines. Additional results are included in the supplementary material.
Number of Refs.4 frames6 frames8 frames10 frames
PSNR ↑\uparrow / SSIM ↑\uparrow31.42 / 0.94231.94 / 0.96031.93 / 0.96131.05 / 0.924
Table 2: Effect of the number of reference images on BlurRef performance on the GoPro dataset.
Figure 5: Visualization of blur and sharp images along with corresponding reference frames used for training BluRef.
DatasetGoProRB2V
Δ=10{\Delta}=1036.1%33.7%
Δ=20{\Delta}=2028.4%25.2%
Table 3: Percentage of matching content between blur and sharp images across the dataset.
ModelsNIQE ↓\downarrow /FID ↓\downarrow
BSRGAN13.34/10.25
Blur2Blur (GoPro as intermediate blur)12.01/8.93
Blur2Blur (RSBlur as intermediate blur)10.07 / 6.28
BluRef10.43/6.45
BluRef + Blur2Blur (RSBlur as intermediate blur)8.47/5.62
Table 4: Performance comparison of unsupervised models on the PhoneCraft dataset. The best scores are in bold and the second best scores are underlined .
Figure 6: Qualitative comparison of BluRef, Blur2Blur, and BSRGAN methods for real-world scenario with PhoneCraft dataset.
Figure 7: PSNR scores (by training iteration) of pseudo-sharp images generated by BluRef when comparing to the ground-truths.
 GoProRB2V
BluRef31.87 / 0.95527.72 / 0.808
Blur2Blur (RSBlur)31.20 / 0.93227.97 / 0.812
BluRef + Blur2Blur (RSBlur)33.30 / 0.96329.62 / 0.872
Supervised - uppperbound33.32 / 0.96228.54 / 0.824
Table 5: Comparison performance between combining BluRef and Blur2Blur with supervised deblurring network. For each test, we report PSNR ↑\uparrow / SSIM ↑\uparrow scores as evaluation metrics.
Annotation for trainingGoProRB2V
Real ground truth33.32 / 0.96228.54 / 0.824
BluRef pseudo-ground truth31.97 / 0.94927.73 / 0.817
Table 6: Comparison between the models trained on original ground-truth data (upperbound) and the models trained on paired datasets generated by BluRef. For each test, we report PSNR ↑\uparrow / SSIM ↑\uparrow scores as evaluation metrics.
NAFNet (light)BSRGAN + NAFNetRSBlur + NAFNet
10.25/6.8513.34/10.2511.4/7.87
Table 7: Comparison between the model trained on the paired dataset generated by BluRef and two generalized image deblurring baselines on the PhoneCraft dataset, using NIQE ↓\downarrow /FID ↓\downarrow metrics.
More in this topic

Image Deblurring

Image deblurring restores sharp images from motion, defocus, and other blur by estimating lost detail and modeling image formation.

See all 62 papers

Explore similar work

CardsList

  1. UniDeblur: Unified Motion Deblurring via Fourier Convolution Learning

    Oct 9, 2026Xintian Mao, Jiansheng Wang, Huaiyuan Wang +2Image DeblurringFrequency-Domain Image Processing

    Shanghai Key Laboratory of Multidimensional Information Processing, East China Normal University, China. · Shanghai Institute of Technical Physics, Chinese Academy of Sciences, China.

  2. Leveraging Phase Information to Boost Unrolled Network Learning for Image Deblurring

    Jun 30, 2026Samira Malek, Haichuan Zhang, Chul Lee +1Image DeblurringFrequency-Domain Image Processing

    Pennsylvania State University, University Park, PA 16802, USA · Dongguk University, Seoul 04620, South Korea

  3. Physen-Noise2Noise: Physics-Guided Self-Supervised Defocus Deblurring with Bias Correction under Low-Light Conditions

    May 23, 2026Ziyan Huang, Lang Wu, Hongji Wang +3Unsupervised Image RestorationImage Deblurring

    School of Mathematics and Statistics, Central South University, Changsha, 410083, People’s Republic of China · Key Laboratory for Micro/Nano Optoelectronic Devices of Ministry of Education, Hunan Provincial Key Laboratory of Low-Dimensional Structural Physics and Devices, School of Physics and Electronics, Hunan University, Changsha 410082, People’s Republic of China