This is just the brute force solution. I'm pretty sure that's no more efficient than guessing all pixels at once and trying to check of that worked or not.
Guessing all pixels at once and then checking is massively worse, since it's basically the first step of GP's proposed approach -- which then iteratively changes things in a way that never makes it worse.
> This is just the brute force solution.
It's the trivial solution. It's significantly more efficient than using backprop. It's much simpler and leads to practically the same result.
This makes me wonder why you would use backprop to do this. It feels like a waste of time, and I'm not sure why this post generates so much attention on HN.
Also see https://nullprogram.com/blog/2020/11/24/
> I'm pretty sure that's no more efficient than guessing all pixels at once and trying to check of that worked or not.
It's iterative so you don't have to run for a very long time, and it will keep improving the loss function for longer.