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..
Degradation-Aware Dynamic Schrödinger Bridge for Unpaired Image Restoration
Authors: Jingjun Yi, Qi Bi, Hao Zheng, Huimin Huang, Yixian Shen, Haolan Zhan, Wei Ji, Yawen Huang, Yuexiang Li, Xian Wu, Yefeng Zheng
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on multiple image degradation tasks show its state-of-the-art performance over the prior arts. Extensive experiments on multiple image restoration tasks, including denoising, deblurring, and dehazing, across a range of degradation conditions, in terms of both visual quality and quantitative metrics. 5 Experiments. Quantitative Evaluation. Table 1 shows that DDSB achieves the best performance across all four unpaired image restoration tasks. DDSB consistently outperforms all competing unpaired methods in both PSNR and SSIM. 5.3 Ablation Studies |
| Researcher Affiliation | Collaboration | Jingjun Yi1,2 , Qi Bi3 , Hao Zheng4 , Huimin Huang4, Yixian Shen3, Haolan Zhan5, Wei Ji6, Yawen Huang1, Yuexiang Li7, Xian Wu4, Yefeng Zheng1 1 Westlake University, China, 2University of Alberta, Canada 3University of Amsterdam, the Netherland, 4Tencent Jarvis Lab, China 5Monash University, Australia, 6Yale University, the United States, 7University of Macau, Macau |
| Pseudocode | No | The paper describes the method and its components (DOT, DTC) using mathematical formulations and textual descriptions, but it does not include a distinct pseudocode or algorithm block. |
| Open Source Code | No | The datasets this paper uses are publicly available, and the source code is promised to be public once published. |
| Open Datasets | Yes | Rain200L [52] is used for deraining. It includes 1,800 synthetic rainy images for training and 200 for testing. Raindrop [34] is used for raindrop removal. It consists of 1,119 paired images with and without raindrops on glass surfaces. The Go Pro dataset [31] is widely used for image deblurring, containing 3,214 high-resolution blurred images (1280 720 pixels). LOL [48] is used for lowlight enhancement. It consists of 500 image pairs captured under normal and low-light conditions. RESIDE provides 13,990 synthetic hazy-clear image pairs in ITS and 500 outdoor test pairs in OTS. NH-HAZE 2 [1] contains 25 image pairs with non-homogeneous haze for more challenging and realistic evaluation. |
| Dataset Splits | Yes | Rain200L [52] is used for deraining. It includes 1,800 synthetic rainy images for training and 200 for testing. The Go Pro dataset [31] is widely used for image deblurring, containing 3,214 high-resolution blurred images (1280 720 pixels). It is split into 2,103 samples for training and 1,111 for testing. LOL [48] is used for lowlight enhancement. It consists of 500 image pairs captured under normal and low-light conditions. Following prior works [20], we use 485 pairs for training and 15 for testing. RESIDE provides 13,990 synthetic hazy-clear image pairs in ITS and 500 outdoor test pairs in OTS. |
| Hardware Specification | No | The time is measured on images of the size of 512 512 pixels using a single GPU. (Table 3 caption). This is too vague ("a single GPU" doesn't specify model). The NeurIPS checklist says "The hardware, especially the GPU requirement, is limited in the subsection of implementation details." but there are no specific details provided in that section. |
| Software Dependencies | No | The paper mentions techniques like MINE and DDGAN, but does not specify any software dependencies with version numbers. |
| Experiment Setup | Yes | The training terminates after 400 epochs, with a batch size of 1. The initial learning rate is 2 10 4 and decays to zero linearly. All the inputted images are firstly resized to 512 512 pixels and then the image intensity is normalized to [ 1, 1]. The time interval [0, 1] is discretized into N=5 uniform steps. The temperature parameter is fixed as τ=0.01, and the balanced constant hyper-parameter of LDOT λ is set to 0.01. For DTC, the weight of degradation term adopts a cosine annealing schedule. |