v3.0: 24k steps trained on a cleaner set of 132 POV videos.

Trained concepts: boobjobs, cumshots, handjobs, vaginal sex (front/back), licking and blowjobs.

Significantly improved dialog and speech issues.

I'm working on NSFW merges on my Huggingface, which will be better than using this LORA alone.

v2.0 doesn't exist, it was a failed run.

v1.0: 16k steps trained on 126 random POV pornographic video/audio data. Used ai-trainer. Rank 64 LORA.

Cumshots, handjobs, vaginal sex, boobjobs, blowjobs and some licking included in the dataset. Overall works pretty good, but not great.

Only use with I2V, which the model was trained with. First to last frame for absolute best results. T2V probably sucks.

I'm working on an actual merge with other good NSFW LORAs. See my workflows and merges here: https://huggingface.co/Phr00t/LTX2-Rapid-Merges

My example was using my "DoEverything" workflow with this LORA at 1.0 strength on top of my "v3 SFW" LTX2 merge.

TRAINING TIPS: Loss graphs are kinda garbage and I didn't really see a drop in error until after ~8k steps, and loss only went from like 1.05 to 0.8, yet results are pretty good. Short videos that focus on a specific action seem to work pretty good with clear prompting of both visuals, motion and audio. Set "caption dropout" to 0 (ai-toolkit defaults to 0.05) and use the "experimental" differential guidance set to 2.5 (at least those are the settings I used). I'd focus far more on video data than image data.
Want to help me pay for runpod time? Paypal donations accepted at phr00t at gmail dot com!