Well, another cool field of activity for AI algorithms in image processing seems to be the RAW development of low-light images. There were already functioning denoising approaches with autoencoders, but the approach now presented in a project goes even further. To illustrate, perhaps first of all the accompanying video:
The trick remains fascinatingly simple: The net gets short exposed RAW shots and learns to derive shots from learning data by long exposure of the same subject. Learning is done with a fairly normal convolution topology. The project data is available on GitHub for you to try out. The network works directly with raw sensor data and replaces a large part of the traditional image processing pipeline, which, according to the authors, only works poorly under extreme low-light conditions anyway.
When you see how much better the recordings of a low-light specialist like the A7SII get with this method, you can be astonished.
;)