Reproduce “Image Denoising and the Generative Accumulation of Photons” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Alexander Krull, Hector Basevi, Benjamin Salmon, Andre Zeug, Franziska Müller, Samuel Tonks, Leela Muppala, Ales Leonardis
The Generative Accumulation of Photons (GAP) framework re-conceptualizes image formation under shot noise as a sequential process of photon arrivals. This approach enables competitive self-supervised MMSE denoising, generates diverse plausible clean images from noisy inputs, and functions as an effective generative model, even when trained exclusively on low-signal-to-noise ratio microscopy data.
Key insights
- Modeling image formation as the sequential accumulation of photons provides a physics-informed framework, where predicting the next photon's location is mathematically equivalent to MMSE denoising.
- The "photon splitting" self-supervised training strategy enables effective denoising and generative modeling directly from noisy inputs, eliminating the requirement for clean ground truth images.
- Iterative photon sampling from the learned distribution facilitates diversity denoising, capturing the inherent uncertainty of the ill-posed problem, and supports full generative synthesis from scratch.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.