It's a Lora I've been training carefully to achieve maximum realism without training real people. I adopted the shapes and poses via Controlnet and other methods to provide a reasonable level of realism, which bothered me with almost all of the loras here. 80% of the loras on Civitai look terrible, as if you're back at Flux1 with a plastic skin and an over-the-top AI look. Flux2klein9b is such a good and mega realistic model, because of all the loras published here the model seems rather bad and looks like it is flux1 as I wrote above. That made me train myself again. The lora is also perfect for image editing; it doesn't change the face, and you can combine it with other loras without major changes. I created my sample photos in 8 steps using Turbo Lora, Flux2klein9b base + Turbo Lora CFG1 + 8 steps, normal Flux2 scheduler with Euler at a resolution of 1024 x 1536. The Lora was trained with natural language, so you don't need any special "triggers." I'm not a fan of pubic hair, so the prompt is usually „shaved vagina“. If you like hair, you need to test if it works :-D.

"The Lora is not finished yet, I will expand the dataset and improve and expand some more. It was just tests at first, but I think it's working well so far, so I've already published it. More version will follow in the next days/weeks."

I have a little tip in case mutations occur in general for Flux2klein9b - All models. I use the "Sigma Power" node from ClownsharkBatwing /RES4LYF and if there are any mutations, I simply change the power to 1.25 - 2.0 or 0.30-0.90 depending on the resolution. This helps a lot with mutations and significantly increases the quality.

v1.1 has more training steps