https://news.developer.nvidia.com/ai-can-no...-grainy-photos/
https://arxiv.org/pdf/1803.04189.pdf

I'd love to see this applied to some famously noisy planetary science images.

If I read this right, the novelty here isn't being the first to try denoising with a deep neural network, but the first to try to "train" the network with only noisy input data. Sounds like several noisy images of the same target train the network how to denoise similar single images. So it's like you could take a bunch of low-light calibration target photos and be able to train a network to then denoise all your low-light photos.