Hello everyone. This model has two built-in LORA models, the datasets for which I spent a long time collecting.

For the anatomy, 532 hand-captions images were used, followed by training at rank 256.

For styles and other variety, 1,000 images with automatic captions were used, followed by training at rank 128.

I can't show all the examples of what the model is capable of; I don't have the time to check every image used for training. I can only say that illustrations, cartoons, and anime weren't used, although I admit the possibility of their transformation.

There are no plans for subsequent versions yet, but if they appear, I won't mindlessly embed LORAs, but will supplement the datasets with subsequent training from scratch.

NSFW sex content is also trained; I don't see the point in showing it, because KREA 2, like other models on the QWEN text encoder, doesn't differentiate between anal and vaginal sex, subsequently memorizing only one thing for each position. To choose the correct orifice, use separately trained LORAs (I have some).