How does the AlignProp option work practically?

[madrooky]

As in the title. I think I understand in principle what it is but it doesn't tell anything how it works, which is quite important to use it to the best benefit. Sure I can simply turn it on and try it out, but the amount of testing and guess work involved in these things can lead to many misunderstandings. And it seems there is no mention about it in the Wiki.
It seems to me it uses the entire data set since there is no option to define specific samples for alignment, that also would mean that the training time would be significantly longer.
A bit of clarification about this option would be super nice.

[madrooky]

Okay I tried it on a small test data set and it turns out to be unusable on my 12GB card, as it loads a clip model for the alignment in which is a bit too much on my vram. Even though the paper about backpropagation alignment suggests a reduction on vram, it turns out to be the opposite in this case.
Although on my settings it should have run somewhat, as I have a setting that only uses 8gb vram and the clip model is a bit less than 4gb. It still uses up everything as soon as the clip model starts to load.
A shame, i though that could be something really cool to play with.

[setothegreat]

Was interested in trying the option out myself, but despite (just barely) fitting within my 24GB VRAM limit, it slows training to an absolute crawl (5 minutes per epoch without vs 30 minutes with). When combined with the rather little documentation this makes it rather prohibitive to test parameters, and doesn't seem particularly worthwhile because of it.

[O-J1]

As in the title. I think I understand in principle what it is but it doesn't tell anything how it works, which is quite important to use it to the best benefit. Sure I can simply turn it on and try it out, but the amount of testing and guess work involved in these things can lead to many misunderstandings. And it seems there is no mention about it in the Wiki. It seems to me it uses the entire data set since there is no option to define specific samples for alignment, that also would mean that the training time would be significantly longer. A bit of clarification about this option would be super nice.

Not a single person in the discord has found satisfactory results with it, myself included, same with Nerogar. He has mentioned he is considering removing it.