V 2.0 is OUT !
New Dataset with a better resolution
Better skin texture and genitals
Full workflow included in the examples.
Feel free to show your creation, really want to see what you can do with it :)
Recommended usage :
Use Acorn Is Spinning checkpoint
weight : 0.7 - 0.8
keywords : 30mm Amateur photo, soft film grain, natural lighting, Natural detailled skin, skin pores
Training :
Trained on 45 images for 8000 steps using a RTX 4090
learning rate : 0.0001
Description
First try making a Lora
FAQ
Comments (12)
Great Job! Could I ask what learning rate you used? It's really picked up the skin texture well.
Thank you :) i used 0.0001
1GB ... can you reduce the rank or reduce to FP16
@sevenof9247 yes ill improve this, first time doing a Lora i'll update it :)
I noticed nud3 in the training, is this intended a replacement trigger for nude/naked
Yeah, it was set as the trigger word but not really useful in my tests
It's my first lora, still a lot of experiments ^^
@aip0pp Good work, I will say in looking through the tokenizer for T5 and Clip-L - T5 is trained with naked, but not nude. CLIP-L does have nude as a token. But it is unclear how they interact. Other trainers have recommended training CLIP-L but not T5.
@Felldude to be honest i don't understand a lot of what you saying, have you any article or link recommandation to learn and improve my knowledge ?
@aip0pp You can find numerous video's from "experts" on youtube, I have over 1000 local trainings and I am far from an expert. I won't knock the people making the videos and I have learned some tips and tricks from said video's but I would always refer to the scientific papers such as this one on the T5 Encoder (Edit Corrected Link)
@Felldude thank you ! will take a look (i recommended your last breast lora, it makes a nice combinaison !)
@aip0pp Thanks, updated the link, as I linked to the wrong thing, you can get an idea from the paper how it breaks down natural language, however the exact interaction is unknown to everyone especially with 2 text encoders. Machine Learning to "AI" where its "AI" because we gave it 2 unknowns and a dropout layer
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.