Trained on 80 hand-picked images with Rank 32 at 512, 768 and 1024 resolution.This time I used JoyCaption, which needed relatively little correction, though correcting 80 images still took a while :PThe dataset included:<ul><li>Blowjobs, handjobs, vaginal sex, anal sex</li><li>One woman with two dicks, two women + one penis, and two women with two different dicks</li><li>Side view, from above, from below, and from behind</li><li>Close-ups and wider shots</li><li>Various rooms and lighting conditions</li></ul>It transfers well to many different ideas and art styles.Enjoy! If you can remember, it would be awesome to see your creations when you tag the LoRA.I used the new validation loss graph in AI-Toolkit to find the point where the LoRA starts to overfit. It works pretty well. I trained up to 6500 steps, but after a lot of testing the sweet spot was around 5000 steps, even though the graph kept decreasing a bit after that. The validation loss is calculated on images that were not part of the training data (I used three different ones).<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/2819b015-4bcc-4576-86ec-73171b1adac9/original=true/2819b015-4bcc-4576-86ec-73171b1adac9.jpeg" />

Files with SHA256: 9a0d6c2c9c540ca6311568b136f571566151510a1c91e9dcf1a7f967dbd3c9e3