This dora is based on my PottingMix PDXL checkpoint.
As this is a DoRA, you'll need need A1111 1.9.0 or newer or apply this PR on your Forge installation.
Sample images were generated using Euler a on AutismMix Confetti.
Recommend using with base PonyDiffusion or AutismMix.
Thanks @mfcg for letting me steal their description template and for helping me with the training 🥰
Feel free to leave me a tip if you like what I am doing :)
Description
FAQ
Comments (12)
This model looks incredible! Could you please share any info on how you trained it??? 💕
basically i collected pics from old animes, captioned it and trained it.
@ElizaPottinger LMAO.
Number of training images? reg images? settings?
@wktra For a simple style I try to get between 75-125 images, adjusting repeats dependingly, no regs. Around 40 epochs, takes me about 6h to train a style.
@ElizaPottinger thank you!
Do you use Kohya?
Onetrainer?
And what's your formula for repeats?
I don't usually set them to anything more than 1.
75 images at 40 epochs overtrains, no?
I appreciate your input because I'm so trying to get the hang of Dora training.
@wktra for 75 i am using 2 repeats, no real formula, just trial and error. Using derrian.
@ElizaPottinger I appreciate you taking the time to answer my questions 💕
Last question: how do you approach your captions?
@wktra For styles? Just let it autotag. Remove some stuff that seems really really wrong and let it be.
@ElizaPottinger thank you so much ðŸ˜
@ElizaPottinger Thanks to your advice, I trained an anime style Dora via PonyXL. Which rank and alpha values did you use?
I also wanted to tell you that I used your advice to train that same style in AnimagineXL and the results are a million times better than with Pony 🤯
@wktra hey, i don't really fiddle much with the settings, so i can't really tell you. Yeah, pony is not the best base :D I hope Pony6.9 is gonna be better :D
@ElizaPottinger I was not a supporter of Pony when it came out. It just seems needlessly complicated in terms of prompting, and I HATE that the underlying base is anthropomorphic ponies. I also HATE that it's one more mid-tier SDXL competitor that requires a whole different set of LORAs to build upon. Just more space to hog on my hard drive.
I finally gave in learned to use it because it did anime aesthetics REALLY well. Maybe 15% better than AnimagineXL. So I began building a LORA collection, alongside SDXL LORAs that did the exact same thing, same characters, same concepts.
And now I'm training with it as a base and it's just dogshit to train.
I'm back at square one with Pony: it's just extra.