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..
Efficient RAW Image Deblurring with Adaptive Frequency Modulation
Authors: Wenlong Jiao, Binglong Li, Wei Shang, Ping Wang, Dongwei Ren
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that Fr ENet surpasses state-of-the-art deblurring methods in RAW image deblurring, achieving significantly better restoration quality while maintaining high efficiency in terms of reduced MACs. Furthermore, Fr ENet s adaptability enables it to be extended to s RGB images, where it delivers comparable or superior performance compared to methods specifically designed for s RGB data. |
| Researcher Affiliation | Academia | 1 School of Mathematics, Tianjin University 2 School of Computer Science and Technology, Harbin Institute of Technology 3 School of Artificial Intelligence, Tianjin University EMAIL EMAIL |
| Pseudocode | No | The paper describes the model architecture, equations for operations within blocks, and flow of data, but does not present a formal 'Pseudocode' or 'Algorithm' block. |
| Open Source Code | Yes | The source code and pre-trained models are publicly available at https://github.com/Wenlong Jiao/Fr ENet. |
| Open Datasets | Yes | We evaluate our method on five datasets: Deblur-RAW [20] in the RAW domain, and Go Pro [27], HIDE [31], Real Blur-R, and Real Blur-J [29] in the s RGB domain. |
| Dataset Splits | No | For the Deblur-RAW [20] dataset, we adopted the preprocessing methodology used by Raw Net [20]. This involved subtracting the black level and subsequently normalizing the raw data to the range [0, 1] by dividing by the maximum signal value. During training, 128 128 patches were randomly cropped from the normalized single-channel raw images. These singlechannel patches, containing the RGGB Bayer pattern, were then packed into a 4-channel format which served as the network input. We employed the Adam optimizer with a batch size of 16. The initial learning rate was set to 0.001 and decayed using a cosine annealing scheduler over 1000 training epochs. All models evaluated on the Deblur-RAW dataset were trained by us on NVIDIA RTX 5880 Ada Generation GPU. For extension to s RGB images, we only changed Cin = 3 in Fr ENet without any other modifications. For the Real Blur [29] dataset, we used the same settings as Deblur-RAW except that training was conducted using 256 256 3 patches. For the Go Pro dataset [27], we adopted training configurations from NAFNet [5]. For evaluation on test sets, we employed the sliding window strategy [8] to process full-resolution images. The sliding window size was equal to the training patch size, and the overlap size was half of the window size. Specifically, for the Real Blur test set, we utilized the official image alignment method provided by the dataset creators [29], ensuring a fair comparison. |
| Hardware Specification | Yes | All models evaluated on the Deblur-RAW dataset were trained by us on NVIDIA RTX 5880 Ada Generation GPU. |
| Software Dependencies | No | The paper mentions using the Adam optimizer and cosine annealing scheduler, but does not specify software versions for these or any other libraries/frameworks like Python, PyTorch, or TensorFlow. |
| Experiment Setup | Yes | We evaluated two model configurations based on the Fr ENet architecture: Fr ENet was configured with a feature width of 32 and 24 processing Fr E-Blocks, and Fr ENet+ employed a feature width of 64 and 20 Fr E-Blocks. Within every AFPM, we divide the feature map into an 8 8 grid of non-overlapping patches. During training, 128 128 patches were randomly cropped from the normalized single-channel raw images. We employed the Adam optimizer with a batch size of 16. The initial learning rate was set to 0.001 and decayed using a cosine annealing scheduler over 1000 training epochs. For the Real Blur [29] dataset, we used the same settings as Deblur-RAW except that training was conducted using 256 256 3 patches. For the Go Pro dataset [27], we adopted training configurations from NAFNet [5]. In terms of the loss function, we used a weighted sum of L1 loss and Frequency Reconstruction (FR) loss Lfr: L = L1 + 0.01Lfr, where Lfr = ||F(ˆI) F(I)||, and ˆI, I, F represent the deblurred image, the ground-truth and FFT operator, respectively. |