DarkStyle版本v1.0 (ID: 374645)

Denizens of this website with the most rapidly worsening UI on the internet!

Do you ever dissassociate so violently you go out for a smoke and come back to a trained LoRA? No? Just me? Okay.

I bring you the results of me coming back from the dead, with a weird thing that came about as a product of my residence within the gin and tonic dimension™, scrawling pinterest for weird things at 2am, and more curiosity about adaptive LR schedulers than sense, I plugged a whole bunch of weird images from pinterest, hit train and went into bed to sober up.

The resultant thing that my poor tortured 3090 spat out, is a lora that does SOMETHING I'd personally put next to the dark fantasy nebulous book cover style.

The training dataset is made up of images with no human faces, and generally, much like me when at the liquor store, operates on v i b e s. This means a refinement is recomended for maximum quality results.

It's trained using an anime model, so, use clip skip 2, you filthy animals.

I've had good results at all sorts of weights, but I would personally suggest around 0.5-0.8, I will suggest you use darpstel as the prompt for the lora, (elaborate atmosphere:1.4) helps too, but it might affect your AI slop result a little too much, so your mileage may vary. The LORA will work a bit without the trigger, so if you are getting too harsh of a stylistic impact, it might serve you well there (I don't know why I have to tell you that here, that is pretty standard practice for LoRA when weight doesn't do it).

As is standard, the model was trained 768*768, so multiples of that resolution are suggested at least in one dimension.

Have fun, post the results.

描述:

训练词语: darpstel, (elaborate atmosphere:1.4)