Twisted-Reality-Krea2
This is a LoRA model with rank 64, trained on a custom dataset that consists of celebrity photo portraits, playboy photos and other NSFW photos.
Concepts
Since this is a rank 64 LoRA model, it is quite huge, containing many concepts such as:
Breasts (gigantic, large, prominent, fake, surgery, augmented)
Vagina (hairy, wet)
Dildos
Penis
Vaginal sex
Usage
Use natural language with this model, and the standard Krea2 turbo model. This is quite a strong LoRA, it deteriorates a bit the image quality, but it brings many hard concepts to the base model. The maximum LoRA strength I would suggest is 1.0, otherwise the image grain gets too destroyed, this is a problem with training such a big model.
Training
Version 1
I trained this LoRA model using 4 bits quantization of the base model and text encoder, using A40s from RunPod. I started setting the learning rate to 2.5e-5 and batch size = 1, 700 steps total, so in total it ran over 700*8=5600 images. I might post an another version with a lower learning rate since the image generation exploded at around 2000 steps.
Version 2
I figured out that doing unguided training on krea 2 destroys the distillation from the turbo model. This version fixes this and there are no more artifacts. This version has been trained on over 56000 images, so it is quite good at generalization, but this won't make the images magically realistic. You will need some other style LoRAs for the realistic skin, etc.
Description
Fixed the training code to not do unguided training. This fixes the degradation of the image quality I was observing in the first version.
lr=2.5e-5
7000 steps
batch size = 8
total images = 56000
FAQ
Comments (2)
ahh, v2 \o/ I was getting some really interesting stuff from v1 even just using it at low strength but gave up after a while cos of the issues. Just did a couple of tests and this is so much better :) Got to go about my day now but can't wait to have a proper play with this later, looks like it might be my new fave lora <3
V1 was broken indeed, my training code wasn't working well with the turbo distillation. Now the LoRA version is pretty much done. I might do a DoRA version in the future...
















