So hello, this is a first post of a finetune on Z image Base
the goal was to able a global understanding of the concept of a woman with a penis.
So I trained it with a mix and varied dataset of 643 images, with various styles, body types and angle and poses, the result is vastly unperfect, but this is a base for more targeted Lora, I would like to offer it to thoses who want to push further on their side, while I am trying to use clever ways to cook a V3
So do not hesitate, download, test and use it as a base for training or merging, would be happy to hear it has been useful.
Fully trained on Z image base, so this is pure, and not yet overtrained, a perfect candidate for further works.
Rank 32, trained in 50 epochs, with a slow degrade of the Lr, and a shift of 1.7 for the last 10 epochs, trained on a H100 with the Musubi script. 12 hours of training.
Description
Better overall, reaching the limit of base models, the concept is cleaner and way more integrated.
The finetune of the Lora is cleaner, with flow shift and Lr decreasing, filtered dataset, while better natural cpationing.
The model keep variety it is not overfit, it change a bit none the less the quality of Z image base.
the anatomy is good in front and side view, but from behind this is hit or miss, and more often time miss.
It retain, realistic, artworks, and various body build, it can understand, flaccid and errect, it have some control on size and shape of the woman, ethnicity is okay inside.
Breasts and nipples are also by accident improved so why not.
Recomended cfg 5, 40 steps, Higher resolution=better results.
this is my last Lora of this kind, I will use the lessons learned to begin a full anatomical model, the pipeline and the technique is good, we need to grow out of the base model.
The next step is about global posing and anatomie, male, female, inbetween, it will be very varied, we will see, wait and see.