Doggystyle POV

Fresh from a break, first epoch I test and... voila!

Versions

One

Pre-vacay I'd already collected 8 clips of this. I quickly captioned it, tailored the buckets to match and trained. I extracted 5 epochs to test, started with epoch 90. It hit and hit and hit some more, with nothing more than the usual bad render here and there. Not sure I'm even going to test the other checkpoints I set aside since this one seems to be working super well.

Let me know if you disagree and why!

Wildcard Prompt

POV doggystyle sex seen from above,

A beautiful {Russian|French|Swedish|Swiss|Austrian|Dutch|English|America|Californian|Siberian} woman with very long {blonde|brown|dirty blonde} hair is having doggystyle sex with a man.

They are on a {white|red|blue|yellow|orange|multicolor|purple|teal|black} {bed|ottoman|couch}.

The {African|Asian|Latin} man's mostly hidden {giant |long |large |}penis remains firmly inserted inside her vagina.

Lighting is {bright and even|soft}.

Training Notes

How I select epochs

Above is the epoch loss curve in Tensorboard. It's taken me a long time to start 'seeing' these properly, the trends, the plateaus, the squiggles...
Note epoch 90, which is at the tail end of a plateau (flattening). I've trained more than 20 LORAs and the sweet spot is almost always somewhere between 0.03 and 0.05 loss. The loss does decrease to closer to 0.02, but that's where I find generalization suffers and overfitting rears it's ugly head!
My training runs always start bumpy for ~10 epochs, decrease steadily for another 20% of the run, plateau a bit for ~10%, then decrease again for another 20-30% of the run. This plateau, before the last drop towards 0.02 is typically where I've found good results in terms of stability and generalization.
For this run I also extracted epochs 65, 75, 95 and 100, but 90 was my first choice because it appeared just before the drop to overfitting (or so I assumed). I still haven't tested the other checkpoints, but again, 90 works so well, why bother?!

Disclaimer

Resper rendsponsibly.