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..

Zero-Shot Blind-Spot Image Denoising via Cross-Scale Non-Local Pixel Refilling

Authors: Qilong Guo, Tianjing Zhang, Zhiyuan Ma, Hui Ji

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on real and synthetic noisy images demonstrate the effectiveness of our approach, which in general outperforms existing zero-shot BSD methods.
Researcher Affiliation Academia Qilong Guo Tianjing Zhang Zhiyuan Ma Hui Ji Department of Mathematics, National University of Singapore, 119076, Singapore EMAIL, EMAIL
Pseudocode No The paper only describes the steps of the method in paragraph text and equations, without a dedicated 'Algorithm' or 'Pseudocode' block or structured code-like formatting.
Open Source Code Yes The code of the proposed method is available on Github2. 2https://github.com/Qilnn Guo/Denoising-NLR
Open Datasets Yes Five datasets, including both real and synthetic ones, are used for benchmarking. (1) SIDD [9], (2) DND [53], (3) FMDD [10], (4) Kodak241 and Mc Master18 [54], (5) fast MRI [55].
Dataset Splits Yes SIDD [9]... We use both the validation set (1,280 images of 256 256) and the benchmark set (1,280 noisy-only images) for evaluation. DND [53]... We use provided 512 512 images for evaluation on their website.
Hardware Specification Yes The proposed model is implemented using Py Torch 2.4.1, CUDA 12.4, and an NVIDIA A6000 GPU.
Software Dependencies Yes The proposed model is implemented using Py Torch 2.4.1, CUDA 12.4, and an NVIDIA A6000 GPU.
Experiment Setup Yes The model is trained using Adam optimizer with an initial learning rate of 10 4, and other parameters of Adam are set to default. The auxiliary image is constructed with patch size N = 8.