This LoRA REQUIRES: Realism Engine v2 (lokr) - because it was trained on top of it and it does not contain its weights merged.

Trigger word: puffy pussy

Attention: Krea2 likes to draw chubby women when it sees 'puffy' in the prompt - to reduce that chance you should specify the physique of the woman (ex: skinny woman)

Downsides:

• Reduces the perceived resolution of the output - likely due to mostly 512x512 training. This is more noticeable on faces. Krea2 doesn't like being trained at low res - apparently. You can mitigate this by reducing the strength of the LoRA though I may train a v2 exclusively at 1024x1024 to fix this problem (if there's demand for it)

About V2:

• Resumed training from V1 at higher res and with higher realistic to anime data ratio.

• Its an attempt at reducing the low-res downside of V1 but I'm still not sure how much it improved in that regard. In some cases, it seems the eyes of characters away from the viewer are slightly more detailed but in other cases its worse.

• Because it was trained for longer it means it diverged further from the Realistic Engine lokr and consequently may not be as good as V1 in scenes involving men.

• My captions were done with JoyCaption a long time ago and it does not mention skin complexion. As a result V2 tends to add a lot more skin complexion than V1 - which may make for more realistic outputs but it may not be what you want. This should be solvable with detailed prompts though.

I consider this LoRA experimental because:

• limited training data (with lots of anime style imgs)

• trained rushedly on batch_size=1 and gradient_accumulation_steps=1 for only 5,700 steps at LR=3e-5 and mostly at 512x512

• first time training on Krea2 - the goal here was not to come out with something very good, rather it was simply for me to test Krea2 at training