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..
Bio-Inspired Image Restoration
Authors: Yuning Cui, Wenqi Ren, Alois Knoll
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experimental results demonstrate that the proposed model achieves state-of-the-art performance across three representative image restoration settings, including four single-degradation tasks on nine datasets, two all-in-one settings, and two composite degradation benchmarks, while maintaining both high computational efficiency and fast inference speed. |
| Researcher Affiliation | Academia | Yuning Cui1,2, Wenqi Ren1,3 , Alois Knoll2 1Shenzhen Campus of Sun Yat-sen University 2Technical University of Munich 3Mo E Key Laboratory of Information Technology |
| Pseudocode | No | The paper describes the architecture and method using textual descriptions and mathematical equations (e.g., Eq. 1-6) and visual diagrams (Figure 3), but it does not contain any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code and pre-trained models are available at https://github.com/c-yn/Bio IR. |
| Open Datasets | Yes | The datasets used in the two all-in-one image restoration settings are summarized in Table 10. The three-task setting includes dehazing, deraining, and denoising, while the five-task setting additionally incorporates motion deblurring and low-light image enhancement for both training and evaluation. Table 10: Datasets used in the experiments for two all-in-one settings. Task Dehazing Deraining Denoising Deblurring Enhancement Training set RESIDE ̒ [103] Rain100L [121] BSD400 [130], WED [131] Go Pro [124] LOLv1 [125] Test set SOTS-Outdoor [103] Rain100L [121] BSD68 [123], Urban100 [127], Kodak24 [128] Go Pro [124] LOLv1 [125] |
| Dataset Splits | Yes | The datasets used in the two all-in-one image restoration settings are summarized in Table 10. The three-task setting includes dehazing, deraining, and denoising, while the five-task setting additionally incorporates motion deblurring and low-light image enhancement for both training and evaluation. Noisy images are synthesized by adding Gaussian noise with σ {15, 25, 50} to clean images. Our dataset configurations closely follow those adopted in previous all-in-one studies [34, 35, 37, 36]. Table 10: Datasets used in the experiments for two all-in-one settings. Task Dehazing Deraining Denoising Deblurring Enhancement Training set RESIDE ̒ [103] Rain100L [121] BSD400 [130], WED [131] Go Pro [124] LOLv1 [125] Test set SOTS-Outdoor [103] Rain100L [121] BSD68 [123], Urban100 [127], Kodak24 [128] Go Pro [124] LOLv1 [125] |
| Hardware Specification | Yes | All experiments are conducted using PyTorch on NVIDIA Tesla A100 80G GPUs. |
| Software Dependencies | No | All experiments are conducted using PyTorch on NVIDIA Tesla A100 80G GPUs. |
| Experiment Setup | Yes | the proposed model is trained using the l1 loss computed in both the spatial and frequency domains, optimized with the Adam optimizer. The initial learning rate is set to 1 × 10−3 and is progressively reduced to 1 × 10−7 using a cosine annealing schedule. Consistent with prior Transformer-based approaches [25], our models are generally trained for 300K iterations. ... The batch size is set to 32, and the model is trained for 100 epochs in the three-task setting and 150 epochs in the five-task setting. The initial learning rate is set to 2 × 10−4. For data augmentation, random horizontal and vertical flips are applied. |