<<<<V2 NOW LIVE>>>
The V1 dataset consisted of 60 high quality videos and was trained with the following settings
60 videos
65 frames each
Trained at 768 resolution for 28 epochs and 1000 steps which took 18 hours on a rented H100
I captioned the dataset with natural language and used terms such as Semen, Ejaculation, Spraying, groin, penis etc
The best results i have achieved was with a CFG of 5 and a strength of 1.5 with a prompt example as:
A nude woman is lying on her back on a bed with her legs spread, a man is positioned between her legs holding his erect penis above her vagina. Ejaculated Semen sprays out of the man's penis and covers her stomach and face with semen. Her mouth is open waiting to recieve the bodyshot of semen. The man is moving his penis back and forth with his hand masturbating. The view is from above, Black hair, blue eyes, bob style hairstyle.
The triggerword is bodyshot but i find it works best near the end of the prompt.
I'm really hoping you can get good results and post them here with your prompting examples.
V2 details......
V2 is released and I'm 100% happier with the outcome.
All of the above still applie and the model is much more flexible now.
Only needs strength of 1.1 and CFG of between 3 to 5.
This has taken so much of my time time so appreciate any Buzz tipped.
Lots of example prompts in the videos.
Description
After my dissapointment with V1 i retrained the dataset with better quality video clips and less variation in positions.
All clips were 512 X 244 X 55 in landscape and portrait but only 5 videos were portrait out of 35 clips.
I spent a lot more time with detailed captions and used the trigger b0dysh0t which is different to v1 where i used bodyshot (this messed up the training i feel)
Again use the same type of language to describe the scene but don't ake them too long and use the example prompts supplied.