This LoRA is intended to produce images of in a woman's mouth, and if prompted (and sometimes when NOT prompted), with the woman on top in a 69 position and a man underneath. The trained in point of view is from between the man's legs.

The Qwen LoRA seems to operate well between weight 0.4 and 0.8 with a noticeable overall bias towards preferring 0.6. (OneTrainer)

The Z-Image version operates well between 0.4 and 0.6 and seems to distort some important angles at, and above, 0.8. (AI-Toolkit)

At the time of this writing I was still producing LoRA that I cast through 3 stages, the first two stages producing a training dataset for the next.

First I choose images, either garnered from web, generating synthetic data or a combination of both. If I'm producing synthetic data then I'm often using an Illustrious model, though sometime a realistic one and, to a lesser extent, NoobAI.

I'll use these initial images to produce a LoRA targeting Chroma-HD, using Chroma 48 as the training base. Then I'll generate a second set of training images using that LoRA, ultimately landing on the target.