This model is quite an improvement over the last iteration. Trained on 288 images with questionable tag quality. The goal was to make a LoRA where you can get an output that doesn't look horribly mutated most of the time even without inpainting or using openpose, and I'd say this very much does that. I used "Anime" as the base model. Miraculously, and I have no idea why, it seems to work both on Anime and Realistic style checkpoints.

It's a difficult pose to train. The legs tend to want to fuse together especially from a head on view. I have a feeling it's because I flipped the images during training so I will update if I can fix it. But I'd say 70% of the time it gives good results.

It seems to work best when you provide a location (on a bed, on the floor etc...). Also turn down the CFG to 3-4 if you're finding lots of weird artifacts.

If you are able to generate something halfway decent please post it here. I have a habit of using quality generations in the next dataset.