Learning Clean Generative Processes from Noisy Measurements
What it is
This paper proposes a method to learn generative models of clean images using only noisy observations by combining Noise2Noise regression with a consistency constraint. It demonstrates that this approach allows for the generation of clean images even when training data is limited to high-noise levels, specifically for non-Gaussian noise types like Poisson and Gamma.
The problem it solves
Standard diffusion models and generative frameworks typically require clean ground truth data for training, which is often unavailable in scientific imaging scenarios where measurements are inherently noisy and the clean signal cannot be directly observed.
What you could use it for
- Training image generation models for photon-counting microscopy where clean ground truth is unavailable
- Synthesizing clean images from synthetic aperture radar data
- Learning denoising and generation pipelines for scientific imaging with specific noise characteristics
Why it matters
Enables the training of high-fidelity generative models for scientific and medical imaging domains where obtaining clean, noise-free ground truth data is physically impossible or prohibitively expensive.
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.