0.1 Alpha: heavier focus on faces and fellatio
0.1 Beta version: More focus on style and posing
Description
FAQ
Comments (8)
YOOOO HOW YOU MAKE IT SO SMALL?? Every lora how pony is about 200mb nowdays, even 600-800mb sometimes
I too am curious about this. How did the size go from 87MB to 30MB?
Some tags in scraped booru datasets are "stranded". They have no connection to other tags or really low connections. Original dataset had 1200 tags and I pruned it to 500 by calculating co-occurence count for every tags (in example set(1,2,3)=set(2,1,3)).
I assumed, rare tags has to have some leverage from popular tags or their content can't be learnt
Also I normalized images with Winsorization
Even if you use a single image with a single tag it will output a 160mb+ file (at least from what I recall from training, unless they changed stuff), I think what most people are missing is that they don't know about the resizing script
@eeeeeqeq Yes and no. If your dataset has 300 images and 5000 unique tags, low dimension LoRA doesn't fit well.
Of course if you desire, you can train underfit model with low dim and high unique tag count.
Can you please try training Shadman! I think you can do it so good
Shadman has a lot of pictures with comic book style panels. That makes cleaning the dataset much harder. I will consider
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.