The LoRA also suffers from the pattern as a result of the fp4 training. I did retrain the dataset, but currently I am in a bit of a dilemma. I cannot get the different poses to work in one or a few LoRAs combined. They train easily each for itself, but if I publish those, it would be a set of 13 (!) LoRAs. So - please leave a comment if you have a proposal how to proceed.<hr />Unmodifed Flux won't do certain poses for women - more or less everything where the legs are not close together. Using a pose controlnet doesn't work either ... it simply refuses. There are LoRAs here to fix that problem, but they are only for nudes, and very explicit ones, too. So I wanted to make p*n which is not p*n - there was no picture in the dataset with uncovered female private body parts, so it is perfectly SFW - and yet a bit inappropriate.The dataset has 120 pictures of several poses, I am going to publish here the first attempt after 16,800 steps. In all pictures, the faces have been masked and it doesn't look overtrained, in contrary not all the poses work with prompting alone, some need a pose controlnet like the sqat and the straddle stretch. In V2 I am going to kick out the crossed legs and focus on the bent sit on a floor which doesn't seem to work.The control word is "spreading legs" but it doesn't do much. You can orient yourself at the prompts of the example images. Not yet working with a prompt are the sitting positions in half butterfly position, side bent sit, straddle stretch and figure 4, albeit sometimes you'll get them if you want something else.

Files with SHA256: 4fedf626876c1bdc3747bcb5e747e1b561b48ddca05c50782145f926391d87d4