/vt/ - Virtual Youtubers

Hi guys have a healthy Okayun. Mogumogu~

Hah, gotcha, I have your attention! It is I, the other LORA Rentry guy here.

Good news everyone! I found a fancy trick in the ninja scrolls, turns out you can basically make a LORA from the difference between a model and its training.
How is this relevant to /vtai/? Easy peasy, you can make a LORA out of HLL/HLL2 and use it with any other model without merging. It's also a much smaller file size. This can be a way to also improve traditional mixes by serving as a testing ground of sorts (so you can gauge how much strength you'd need to merge and keep details).

The magickal incantation is something like
"python networks/extract_lora_from_models.py --model_org ../stable-diffusion-models/Stable-diffusion/animefinal-full-pruned.ckpt --model_tuned ../stable-diffusion-models/Stable-diffusion/hll2-e12.safetensors --save_to hll2-e12.LORA.safetensors --dim 192"
Resulting file is a 432MB .safetensors file. I plugged the resulting LORA, generated with AbyssOrangeMix2 et voilá, works very close to the mix.
If someone tells me a decent file storage that can handle files that big and doesn't require me to make an account I'll upload it to save you the hassle.

I'll be putting more exact instructions in the rentry in the next ten minutes: https://rentry.org/59xed3

I can feel the disk space savings already.