Trained on 160 5-15 second clips of women kissing passionately. Trained specifically on lesbian videos, so no idea if it will work for guy girl, guy guy. I haven't tried.
Still iterating, definitely needs more diverse data.
The subjects in the training videos were picked up in the lora a bit - so if you see the same people in T2V over and over lower the strength to .7 or so.
Description
FAQ
Comments (17)
From the samples it looks pretty good. Nice job
Looks great. Is there a way to convert a lora to a lower rank? Or does it have to be retrained? The lora results are fantastic, but the size requirement per concept is pretty big, and I was wondering if I could get similar results from a lower rank.
yes you should locally be able to lower rank. comfy may even have a vanilla node
https://civitai.com/articles/10659
This one resized to 190mb, 95.64%.
@firemanbrakeneck Excited to try this out. Thank you!
"Still iterating, definitely needs more diverse data." You can use regularization datasets to fight this.
what do you use for regs non NSFW data or just different types of NSFW videos
@MisticRain69 I use images of random women that are not in the dataset. I have a dataset that consists of about 1000 images of random women of varying ages and races. I use the caption "a woman." in the reg set. I use the same number of images in the reg set as the dataset.
Does this work only with Sulphur model?
In my case, it worked well on all models (default distilled 1.1, eros v1, sulphur).
Shouldn't this stuff already be in the base model?
can this make bald women?
would love if it had more guys, especially like, a bearded man. I think it's the type of lora that will really improve with a diverse dataset
Thanks, finally a kiss, Lora, it's a shame there aren't any men there, but it still works.
It works with mens in I2V but i did not try with T2V