New method learns clean images from noisy data using consistency relation

Researchers have developed a method to learn generative models of clean images from noisy observations, particularly for diffusion processes and non-Gaussian noise. The technique utilizes Noise2Noise regression to learn predictions from independent noisy views, but requires consistency to generate images below the training data's noise level. This consistency relation ensures that a prediction equals its expectation after continuing the sampling process, which has been verified for various distributions like Poisson and binomial. Experiments on CIFAR-10, microscopy, and synthetic aperture radar data demonstrate the effectiveness of this approach in learning posterior predictions from noisy views alone and repairing continuations beyond the data's noise level. AI

IMPACT This research could improve generative model performance in noisy environments, potentially impacting fields like medical imaging and remote sensing.

RANK_REASON This is a research paper detailing a new method for generative modeling. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• binomial
• CatalyzeX
• CIFAR-10
• DagsHub
• Gamma
• Gotit.pub
• Hugging Face
• Influence Flower
• negative binomial distribution
• Noise2Noise
• photon-counting microscopy
• Poisson
• ScienceCast
• synthetic aperture radar

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →