v0.2

Dataset increased to 55 images with two new poses. I tuned my settings in AI-Toolkit and can now train at Rank 16 on 512×512 and 768×768. Rank 32 should also be possible in the future.

I also improved the quality of all images in the dataset and refined the captions.

Poses in the dataset (the names dont really do anything unfortunately, you´ll need to describe them and don´t describe what would not be visible in that positon):

• lying on her back (26 images, 2 with the second man squatting above instead of underneath her)

• (reverse) king of clubs position

• doggystyle

• cowgirl

• while standing (man behind lifitng her)

• double front position (two guys standing in front of her with their hips close together)

Initial findings:

• Much better image quality

• More reliable results

• The cowgirl position kind of works now

• Stronger than before, 1 is almost too much

• It transfers to art styles

v0.1

My first try at making a LoRA. I'm happy for any feedback, hints or suggestions. And please remember to tag it — I'd love to see your results!

Trained on 24 captioned images. The dataset contained roughly five different poses: on back, regular cowgirl, doggystyle, king of clubs (just found that name lol, didn't use it in the captions yet) and reverse king of clubs.

All of them except regular cowgirl work quite well. That said, you still get better regular double penetration results in the cowgirl position when the LoRA is enabled.

I also found that it is more consistent at 1 megapixel resolution. At 2 MP I needed more seeds to get a good result. Probably because of the low training resolution (I only have 16 GB VRAM).

Training settings:

Rank: 8

Steps: 2250 (results got worse after that)

Resolution: 512x512

Trigger word: double_anal

Recommended strength: 1