Tried to train a lora for the tf2 cast with about 240 curated screenshots from the "Meet the team" videos, got this after some tweaking. Not a good modelbut it kinda works more often than not so i'm leaving this here along with the training dataset until i have some time to improve this in a v2.
V2:
trained at 768x512
V1:
Seems to work best at a resolution of 1024x576 (i blame the image resizing script) and the settings i use in the example images here. Trained via tagging style with:
-tf2mercs
-tf2scout, tf2soldier, tf2pyro, tf2demoman, tf2heavy, tf2engineer, tf2medic, tf2sniper, tf2spy
-red, blu
Description
it's a LoHa now
trained with better parameters
still the same dataset
still needs improvement, especially for multiple mercs in one prompt
FAQ
Comments (6)
we are close brother, we can achieve sfm shitpost quality
There's already some nice shitpost potential with the model just mixing different characters sometimes imo, can't wait to get this right eventually! :D
@BelgOwned ill be the first to download it when it finally becomes perfect
i think you should train the mercs individually
like a lora for each character
so that it can get really close to their designs without any mixup or bugs
train them on a lot of 2d art and fan art to make the lora understand more and less 3d and sfm
since majority of art will be used in anime 2d models
I could definitely try to use the same dataset to train each character seperately, too busy to make a better dataset atm.
@BelgOwned i wish i could send a gigachad gif
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.