Type | |
Stats | 1,947 |
Reviews | (300) |
Published | Aug 8, 2023 |
Base Model | |
Hash | AutoV2 8BE8A4378F |
Usually my LoRA testing ground, someone asked for it so i'll just post it here.
Not the easiest or best model to work with but hey its here if you want to try it.
More examples are in images section using wildcards, randomized prompts and my own LoRA's.
RECOMMENDATIONS :
Clip Skip : 2
Min Resolution : 512x768
Hires-Fix : 1.5-2x
Upscaler : 4xUltraSharp
Denoising : 0.3-0.5
ENSD : 31337
Sampler : Upto you, but i use Euler a mostly
For Better Faces/Eyes : install & use After Detailer
Min Steps : 20
CFG : 5-9
Sample Negative Prompt :
verybadimagenegative_v1.3, bad_prompt_version2, easynegative, bad-hands-5,
For Higher Res Image generation on weaker hardware : enable xFormers, lowvram, use Tiled Diffusion & VAE, set token merging ratio to 0.2-0.3.
Also Dynamic Thresholding for better details
My general upscaling method :
512x768 txt2img, 2x hires-fix with 4xUltrashap, 0.4 Denoising -> send to img2img ->
ControlNet Tile Upscale with Tiled Diffusion & VAE.