RainehDaze

I suppose that's doable, assuming that a given config can't be just set to "Ignore waves" somehow.

I mean, I think that would be the expected outcome? A configuration that only cares about the number of turns processed would either always fire as if the conditions are...

Vulkaninfo works: [vinfo_out.txt](https://github.com/ValveSoftware/Proton/files/2849572/vinfo_out.txt) DXVK normally works (e.g. both Monster Hunter: World and Dark Souls III have no such graphical problems)

I would try it with fewer tokens and no initialisation text. Could also try lower CFG. What might be an issue is the fact reconstruction is broken and there's been...

Hm, I wonder if it's the lack of reconstruction? I checked the paper, and it was pretty clear--three positive tokens, three negative tokens, 0.0025 learning rate, and a mysterious gamma...

 The new img size is under your max size and it's generating a larger image then downscaling, I think.

Looking at the numbers, it might be that your image dimensions have to be divisible by 8 and it rounded down to the nearest size it _can_ generate.

Yeah, 3090 here. I was baffled--had it working a couple of days ago, decided to do it again with a _way_ bigger image set, consistently got nothing. Was wondering "do...

I don't think downgrading CUDA actually helps; your textual inversion still won't learn anything in particular. Some very, _very_ broad elements of a picture, not the intended ones.

As I said, I don't think that the above fix actually fixes it. Consider this prompt matrix:  The data set, training time, and other settings are equivalent. Yet, clearly,...