Style lora trained on art of Vsevolod Ivanov
For some reason I misremembered the artists' surname so the trigger word mistakenly considers him to be Vsevolod Volkov instead of Vsevolod Ivanov, just use the trigger as set in the lora (v4-zimage). My bad.
Description
upped rank to 48
added some more paintings to the dataset
additionally trained on cropped 512x512px tiles of larger paintings which might increase detail
cogvlm captions for paintings
might be a bit overfit at 5000 steps, lowering the weight gives better results sometimes
FAQ
Comments (11)
Its time to restore Great Slavic Grad
I rememebr this artist. But he was good due to small details. Is flux able to recreate this?
Maybe zooming on his art and using parts of it fro trianing would work best?
it's a good idea, cutting some tiles and training on those too, for v1 i just got some random stuff off google images which was mostly low-res.
@randomkcoala3 man. I have folder of his art, and of other slavic art. do you want me to share it with you?
@barlogspam176 sure, i'd love it, although i'm not sure if adding more paintings to the dataset would do much good at this point, there's 73 pics now (and ~130 tiles) and more is not always better with loras.
@randomkcoala3 yeah probably not. My lORA of people done with 30 images are better then the ones done with 131
https://mega.nz/file/0FQXRBIB#RCpJNMtKiTNBFc7btfbPH2fQ0DedCSQDe7ffkAlgqP0
Most fo relevant pics are in the vsevolod folder
@barlogspam176 thx mate, much appreciated.
I didn't know this artist, thanks for sharing!