Gabriel Peyre has entered the chat

I actually tried with this code! No, it didn't work. My explanations are that 1) the forward walk noise has a certain character that is part of the learning, and 2) what I see is that other data does move, but it just moves in tandem too much and the points don't split up enough to cover the whole area. I suppose it fails because close points move together. (As for my 2), one should on the other hand also not try to view the points as individuals.)

It's not like I got a list of strong references - I'm trying to dig for them myself right now. One recent pdf I came across is "Tutorial on Diffusion Models for Imaging and Vision" by Stanley Chan - maybe there's merit to it. If you know the math of the Ornstein-Uhlenbeck process, , which has independent interest and so is also covered e.g. on math books on stochastic differential equations, then the ML part might mostly be plug and chuck and wondering what exactly it is that makes the learnability possible. I feel a lot of people with that topic in particular tie in to physics, so if you've ween the Langevin equations that might help too. But while I love to get lost in theory, with this stuff I feel is it's good to just hack into a python script for starters, like I did here. Hope that helps. PS let me know if you find something thorough too.