Welcome to KayWaii XL!
Although the name was taken from my previous SD 1.5 model, this model is not a complete reboot of KayWaii, it’s more like an attempt to recreate it with the current resources available. This model is built on top of AutismMix Confetti (I know I’m more known as an Animagine XL creator lately and as a part of the actual Animagine team, but I want to try something different.)
Changelog
v1.2: More throughout dataset cleaning, re-tagged dataset with better tagger, improve LoRA capability.
v1.1: Updating dataset and improving training settings so the maximum epoch is usable.
v1.0: Initial release
How to use the model
Prompting
Use this prompt format to help you get better results:
1girl/1boy, (describe what you want, using danbooru tags is recommended), score_9, score_8_up,Use this negative prompt format:
score_4_up, score_5_up, extra digits, fewer digits, (things you want to remove, using danbooru tags is recommended)Settings
Euler a
Steps 25-28
CFG 5-7
The sweet spot is around 25 steps with CFG 5.
Training configurations
Version: v1.2
Training method: LoRA training, then merging the trained LoRA into AutismMix Confetti.
Dataset size: 28,546 images.
Epoch: 10
(I will not give more details in public, since both Pony base and AutismMix did not care to give training details too. But feel free to DM me to ask something about it.)
PonyDiffusion XL v6 and its derivatives fall under the Fair AI Public License 1.0-SD license. (Hey astralite, you should publish training details you know if you’re using this license.)
===
Description
Dataset update, training configuration update.
FAQ
Comments (6)
Accidentally put v1.1 as SDXL (welp Pony IS actually SDXL). Has been fixed and now you can use it with Pony LoRAs on on-site gen.
I'm curious about this model, I feel what Pony needs mostly is more characters and Art Styles to play around with. Still, this is a great first step, at least it has a closer anime look to it. I will definitely need to play around with it later.
Honestly I wanna try that if I have infinite funds. But I don't have that so here I am.
@kayfahaarukku Yeah, completely understandable, still your work is appreciated, hopefully you can pull it out one day!
BTW there is a model that does have a lot of artist and characters, but I feel the results aren't that good, dunno if you have check out 4th Tail model {its Pony}, maybe you can use it as a base or starting point and improve it?
Does this checkpoint need the sdxl vae like base pony?
The VAE is baked into the model so you don't have to use a separate VAE file.