Bit of a redundant LoRA because SDXL seems to know Disco Elysium as a token!

This LoRA is more of an experiment than anything, try out just the Disco Elysium token first.

Nevertheless I had some issues with just using the token, it has a tendency to try and give me isometric top-down view images and a lot of background people, so I trained a LoRA on only the official portraits and wallpapers.

The token I trained it with is dElysium, you can also add the Disco Elysium token.

Trained on SDXL0.9, 1024 res with gradient checkpointing and batch size 2, 8 dim 4 alpha.

This peaked at about 9gb VRAM used (not training the text encoders).

I am using this cursed Comfy workflow https://pastebin.com/raw/marWYmFe.