Remember Disco Diffusion? Well I do, and I loved it. This LoRA is trained on selected images from my own minor successes and major fails with Disco Diffusion. The LoRA behaves very much like DD did, so this is for anyone who's a fan of:
Unpredictability
Incohesion
Overall weirdness
All my DD LoRAs were initially trained for SD XL 0.9 and then retrained for 1.0, I guess in July 2023. I can't remember what settings I used, but very likely they are all over- or under-trained with a poorly captioned limited dataset. Due to the dataset containing quite a few portraits the LoRA probably skews towards faces, and will most likely also affect facial features and expressions.
Elaborate prompts will most likely override the style. Start with something simple and go from there. Adding "aiai-DiscoFaces" to the prompt will enhance the style. There are more trigger words, but I've forgotten most of them; do experiment and see what works. Most likely the word "alien" will do something.
For best (aka more chaotic results) I suggest you use a mix of all my DD LoRAs at lower weights, like 0.3 each or something along those lines. Either way you probably never want to go above 0.75.
Merging these with good checkpoints may also be a decent idea.
All example images are generated in Automatic111 using:
SD XL 1.0 checkpoint
DPM++ 3M SDE Karras
30 steps
CFG 7
1024x1024
Hires fix 4x-UltraSharp denoise ~0.35
No refiner, extensions or other shenanigans.
More Disco LoRAs:
Good luck (you'll need it :)