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 Denoising via Neural Compression: Theoretical and algorithmic framework

Authors: Ali Zafari, Xi Chen, Shirin Jalali

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through extensive experiments, we show that ZS-NCD achieves state-of-the-art performance among zero-shot denoisers for both Gaussian and Poisson noise, and generalizes well to both natural and non-natural images. Additionally, we provide new finite-sample theoretical results that characterize upper bounds on the achievable reconstruction error of general maximum-likelihood compressionbased denoisers.
Researcher Affiliation Academia Rutgers University New Brunswick, NJ, USA
Pseudocode Yes Algorithm 1 Finding Lagrangian coefficient λ
Open Source Code Yes Our code is available at: https://github.com/Computational-Imaging-RU/ZS-NCDenoiser.
Open Datasets Yes We evaluate on grayscale Set11 [25], RGB Set13 [54] (center-cropped to 192 × 192) and Kodak24 [55] datasets. ... we test ZS-NCD on Mouse Nuclei fluorescence microscopy images [56] ... We also assess real-world denoising using the Poly U dataset [57]
Dataset Splits No The paper mentions preprocessing steps like cropping images (e.g., 'center-cropped to 192 × 192' for Set13, 'Images are cropped into 128 × 128' for Mouse Nuclei, 'Images are cropped into size of 512 × 512' for Poly U). However, it does not specify explicit train/test/validation splits for these datasets. Given the zero-shot denoising context, the method trains on a single noisy image rather than a split dataset.
Hardware Specification Yes All the experiments were run on Nvidia RTX 6000 Ada with 48 GB memory.
Software Dependencies No The paper mentions using 'Adam optimizer' and 'GDN [32] for Conv network and Re LU for MLP' as activation functions, but does not provide specific version numbers for any key software components or libraries like Python, PyTorch, or CUDA.
Experiment Setup Yes Adam optimizer is used for training the networks over 20K steps, with initial learning rate of 5 × 10−3 decreased to 5 × 10−4 after 16K steps for the Conv-based network. The learning rate for MLP-based networks is 1 × 10−3. ... For noise levels (15, 25, 50) we set λ = (300, 850, 3000). Similar to Kodak and other experiments we set training epochs to have 20K steps of gradient back propagation. For Poisson denoising α = (15, 25, 50) the λ = (3000, 1500, 1000).