I'm lazy af so I'll just copy paste the details. They are basically the same for all of my LoRA.
This is a LoCon / LoRA model of beautiful Korean girl face. Model was trained on various irl photos, however it does not depict any real person. Consider it a virtual character.
Trained on chillout. Offset noise has been applied.
-
Get this extension: MultiDiffusion with Tiled VAE
Use Tiled VAE only.
Set encoder/decoder tile size to what your gpu can handle, and get faster image generation speed. This extension also allows you to upscale image to much higher resolution, beyond your gpu normally could handle. No more OOM errors!
I have 1080Ti with 11Gb VRAM, here's my setting for Upscale by 2.
For Upscale by 3 (which was impossible for 1080Ti to handle before), I change tile size to be smaller.
base res is 512*768
Sampler: Euler a / DPM++ SDE Karras
Enable Hires. fix
upscale by 2 or higher.
Hires step: 8 / 12 / 18
Upscaler: R-ESRGAN 4x+
Denoising strength: 0.1 / 0.2 / 0.3
-
Enjoy.