Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
BlurDM: A Blur Diffusion Model for Image Deblurring
Authors: Jin-Ting He, Fu-Jen Tsai, Yan-Tsung Peng, Min-Hung Chen, Chia-Wen Lin, Yen-Yu Lin
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that Blur DM significantly and consistently enhances existing deblurring methods on four benchmark datasets. The project page is available at https://jin-ting-he.github.io/Blur DM/. 1 Introduction Camera shake or moving objects frequently introduce unwanted blur artifacts in captured images, severely degrading image quality and hindering downstream vision applications, such as object detection [11, 41], semantic segmentation [1, 43], and face recognition [12, 23]. Dynamic scene image deblurring aims to restore sharp details from a single blurred image, a highly ill-posed problem due to the directional and non-uniform nature of blur. ... 4 Experiments 4.1 Experimental Setup 4.2 Experimental Results 4.3 Ablation studies |
| Researcher Affiliation | Collaboration | Jin-Ting He1 Fu-Jen Tsai2 Yan-Tsung Peng3 Min-Hung Chen4 Chia-Wen Lin2 Yen-Yu Lin1 1National Yang Ming Chiao Tung University 2National Tsing Hua University 3National Chengchi University 4NVIDIA |
| Pseudocode | No | The paper provides mathematical formulations for the dual noise and blur diffusion process, and the dual denoising and deblurring process (e.g., Section 3.1, 3.2, and Appendix A.1-A.3), but it does not include any explicitly labeled "Pseudocode" or "Algorithm" blocks with structured, code-like steps. |
| Open Source Code | Yes | The project page is available at https://jin-ting-he.github.io/Blur DM/. Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: We provide our code in supplementary material. |
| Open Datasets | Yes | Deblurring Models and Datasets. We adopt four prominent deblurring models, including MIMOUNet [5], Stripformer [40], FFTformer [13], and Lo Former [21], to validate the effectiveness of Blur DM. Following previous work [5, 13, 21, 40], we adopt the widely used Go Pro [24] and HIDE [36] datasets. The Go Pro dataset contains 2, 103 image pairs for training and 1, 111 image pairs for testing, while the HIDE dataset contains 2, 025 image pairs used only for testing. Additionally, we utilize the real-world Real Blur [31] dataset, which contains Real Blur-J and Real Blur-R subsets. |
| Dataset Splits | Yes | The Go Pro dataset contains 2, 103 image pairs for training and 1, 111 image pairs for testing, while the HIDE dataset contains 2, 025 image pairs used only for testing. Additionally, we utilize the real-world Real Blur [31] dataset, which contains Real Blur-J and Real Blur-R subsets. Each subset contains 3, 758 training pairs and 980 testing pairs, with Real Blur-J in JPEG and Real Blur-R in Raw format. |
| Hardware Specification | Yes | Analysis of Blur DM s Computational Overhead. We present the computational overhead introduced by Blur DM in Tab. 4, measuring FLOPs and inference time on a 256 256 image using an NVIDIA Ge Force RTX 3090. |
| Software Dependencies | No | The paper mentions utilizing various deblurring models (MIMO-UNet, Stripformer, FFTformer, Lo Former) and optimizing the framework, but it does not explicitly list specific software dependencies like Python, PyTorch, or CUDA with their version numbers. |
| Experiment Setup | Yes | Implementation Details. Fig. 3 illustrates the architectural design of the four components in Blur DM: the Sharp Encoder (SE), Blur Encoder (BE), Blur DM, and Prior Fusion Module (PFM). Specifically, SE and BE have the same network architecture, each with six residual blocks, four CNN layers, and two MLP layers. Blur DM contains noise and blur residual estimators, each comprising six MLP layers. PFM consists of one MLP layer. We empirically set T = 5 in Blur DM, with β1:T increasing uniformly from 0 to 0.02 and α0:T increasing uniformly from 0 to 1. The overall framework (Third Stage) is optimized using the default training settings of each deblurring model, including learning rate, number of epochs, batch size, optimizer, etc., to ensure fair comparisons. ...All experiments are conducted with 1, 000 training epochs used. ...Table 7: Comparison of training cost and performance between baseline and Blur DM. Method Training epoch Training time [h] PSNR [d B] Baseline 1 3000 Blur DM 3 3000 (Stage 1) + 500 (Stage 2) + 3000 (Stage 3) |