You'd assume that a LORA that just shows a pretty face rocking forward and back would be easy! Apparently not. Occlusion is a b!tch.

Front Doggystyle

A{ slim| fit|n athletic| curvy| heavyset| cute} {Russian|French|Swedish|Swiss|Latina|Austrian|German|Dutch|English|Irish|Portuguese|Indian|Russian|Swiss|Swedish|Danish|Italian|Asian|Japanese} woman remains bent forward from the front with a man kneeling low centered behind her, thrusting steadily.

{eye-level face prominent|slightly below eye-level|slight high angle|high angle  showing her butt crack}

Her {small breasts jiggle beneath her|medium-sized breasts swing with each thrust|large breasts swing beneath her}.

{She is very happy|Her mouth is open and her brow is furrowed|She is screaming from pleasure}.

{His right hand is holding her hair|His hand is holding her bangs|His right hand is holding her ponytail tightly on the top right of her head|He is holding her shoulders firmly}.

They are both naked with the scene lit {brightly|softly|dimly with a spotlight on the subjects}.

Versions

three.1

Same as below but a tiny bit cleaner and slightly better(faster) motion.

three

Blurred faces and retrained with a select set, pruning most deviations from the core pose. This allows for stability through generalization. 10 videos, 24fps, 48 frames, ~256x256 LR 5e-5

two

19 videos, 24fps for 50 frames, better captioning and better diversity. We don't need the pillow anymore as the normal resolution run caught the right details down there, most of the time. This one is head and shoulders above pillow edition but far from perfect. Alas, it's worth the release!

pillow edition

This is just a proof of concept I ran. The curve looked great till 140, best epoch is 55... figures...

** This version has some issues with the stomach, so I prompt for a pillow there. A higher rez and lower LR will probably do better for the next version.

Training

Want to contribute? Check out my Training article!

Disclaimer

AI doesn't know any better, but we do.