Tested this idea out. Its difficult to get a perfect result but with in painting you can improve it. Its a silly fun lora.
Update:Attempted a v2 using text as a trigger word. Still didn't nail the result but thought I'd release it anyways for people who enjoy this.
My Links: discord, twitter/ig
Support my efforts at @ patreon
Check IG for more content like this below.
View my article on checkpoints - here. (HUGE ANIME UPDATE)
Image Gen series part 3 is out now
Description
FAQ
Comments (6)
interesting point is that it got the font down, but not the words. There is a chance for a paper how to train a AI to return the correct text.
... and I can scream?
"I have no tits and I must jiggle" by Harlan Ellison
I feel like text might be the better trigger word because anime checkpoints already know what text looks like, so if you train the keywords on the token text, it should regcognize the exact text in each image better.
For all we know, the lora might be associating the text with the keyword black shirt making it all jumbled. Just a theory.
generates everything except the tshirt with "i have no tits" written on it lmao
>nabrnon
>hon ıon
>nne ve ntit ℌ