Model trained on GITS 1995. Use 0.2-1
Description
Trained on Illustrious 0.1. Double the dataset and scaled it to 4k. Characters trained with token warmup. Check example images.
kusanagi motoko \(gits 1995\), project 2501 \(gits 1995\), batou \(gits 1995\),FAQ
Comments (5)
love it! thank you <3
It's almost weird how well Z-Image picks up on the style considering how much less anime training it has than Illustrious. Maybe that's precisely why it works so well?
Could be. More parameters. VAE that isn't complete trash.
@xixxix Yeah, but also training only available through hacks that don't work optimally, on a distilled model. If it's this impressive given that, then the Omni base model is going to be crazy.
Z-Image can do decent anime in a variety of styles just via prompting, so a lora just needs to pull that together and make it more consistent. (This is true of a lots of other things too; many ZIT loras so all the heavy lifting is in the prompt and the lora just finishes things off)