Yes, though it takes a super long time

@bobbysmithy55555426 Sure, I'm happy to share.

The dataset for this specific Lora is 22 clips of ladies doing their thing: 14 clips are real woman and 8 are anime*. All are roughly 5 seconds long, and they are all trimmed to 256x256 pixels. There are probably professional tools that can trim videos, but I do it the caveman way by playing the clip and using the builtin Windows screen capture tool to capture a 256 square section of the screen. (This also ensures that the clips are all of the same framerate).

They are then captioned. I haven't found a good AI tool that will accurately caption nsfw stuff yet so I do this manually. It sucks, but is very important. You describe who is doing what, where. I'll post an example in the next comment.

I then use diffusion-pipe to train the actual lora. I mostly use the default settings in the WAN low-noise config.

I've experimented a little and I've gotten the best results by doing a two-step training, initially at a lower learning rate then cranking it up for the last ~40% of training. For high noise I'll do about 1200 steps at learning rate 2e-5, then about 800 more at 2e-4. For low noise, roughly 1800 steps at 3e-4, then 1200-1600 more at 3e-5. I train locally with my single consumer-grade GPU, so these trainings altogether take like 12-15 hours.

I'm sure there are more professional setups, I'm just a guy with a 3090 and too much free time.

*I particularly like to make anime gens so I try to make sure they're represented well in the dataset, but depending on the act being performed it can be hard to find uncensored clips. For example the handy-bj dataset had none since anime penises are basically always censored.

@bobbysmithy55555426 An example caption for one of the clips in this dataset:

```
The video shows an Asian woman with long black hair. She is completely naked.

She is bouncing up and down, having cowgirl sex with a man, her large breasts bouncing. The view is POV. His penis is sliding in and out of her vagina. The woman is aggressively moving her hips up and down, slamming the penis inside her pussy. The penis goes completely inside the pussy, then comes back out. They are having fast, rough, intense sex.

The video focuses on her pussy, breasts, and face. As the video progresses, she covers her mouth with her right hand.

The background has bookshelves lined with books, and a wall with many posters and pictures, indicating that the scene takes place inside an apartment.
```

I am very lazy, so for the most part I keep the description of the action the same for everything, and just quickly change the description of the actors and the background for each, as well as anything "unusual" that happens (like in the above where she covers her mouth with her hand).

@mrdion It works well for 480p, my GPU can't handle 720p so I haven't been able to test it.

@emilkwok
1. Honestly it's all vibes. The first time I did it I trained at a higher rate for a first try, but it didn't turn out the way I wanted so I tweaked the dataset and (because I am lazy and didn't want to retrain from 0) I tried resuming from about halfway with a higher learning rate to "offset" the old data and it turned out well. So I just keep doing that and it seems to work. I have absolutely no scientific or mathematical reason for doing it this way, maybe it's worse than just leaving the rate the same the whole time :)
2. That's exactly right. I save a checkpoint every 10 epochs, and I manually restart training with a --resume_from_checkpoint command and an updated settings.toml