Quick Tips for PonyXL [Best settings] - Ai Image Generating (Patreon)
Content
PonyXL seems to be kinda overwhelming.
In the last weeks, I spent most of my time going through all models of character, art-styles, and utility,... if you go to civitai.com you can find nearly 100,000 Ai-models for PonyXL... well I cannot test them all but!
THERE ARE THE BEST MAIN MODEL I FOUND SO FOR.
Checkpoints (main model):
3x3mixXLtypeB - https://civitai.com/models/431350/3x3mixxltypeb?modelVersionId=480549
+ Very fast (for rescaling as well)
+ half realistic
+ very good NSFW
+ characters look naturally older
- VAE will break it (don't use vae)
- maybe too realistic
- not every character model responds well
Ceii Mix [PDXL] v1.0 - https://civitai.com/models/482488?modelVersionId=536595
+ perfect balance anime-realism
+ perfect NSFW
- not the best responses for non-nude poses (very stationary)
SakuraSushi XL - https://civitai.com/models/488884/sakurasushi-xl?modelVersionId=543611
+ anime style
+ Every tested character looked well
+ very good nsfw
- very slow
- crashes when 2x+ rescaling!
- not realistic
- hard to change art-style (utility/tool model doesn't work much)
- faded colors (VAE will fix it) if not -> time for Photoshop
Mala Anime Mix NSFW PonyXL - https://civitai.com/models/442163/mala-anime-mix-nsfw-ponyxl?modelVersionId=550933
+ anime style
+ very good nsfw
+ art-style changer based on model (for me it's good, for someone may not be)
- characters look younger
- not realistic
The standard resolution that gives me the best results:
616 x 616
536 x 688
560 x 656
Rescaling settings:
Upscaler --- R-ESRGAN 4x+
Upscale by --- 2.3x
Hires steps --- 12 (less than 12 = bad hands!)
Denoising strength --- 0.35 - 0.5 (the further character is bigger denoise you need)
Nvidia RTX 4070 / 4070ti / 3090 series: speed of generation
version of SD ---- 616 x 616 image ---- 2.3x-rescale
in SD1.5 ---- 3s ---- 10s
PonySL ---- 7s ---- 15-90s (depends on the checkpoint, somehow)
You probably know this page about speed of GPUs: https://www.videocardbenchmark.net/high_end_gpus.html
IT IS NOT ACCURATE! The numbers are for real-time calculation (for playing games,...) for AI-generating and 3D rendering it's completely different.
Look here: https://www.tomshardware.com/pc-components/gpus/stable-diffusion-benchmarks
or here https://opendata.blender.org/benchmarks/query/?group_by=device_name&blender_version=4.0.0
Simply bigger RAM faster gen.