Hey there, I am glad you like my models and thanks for being interested.
First and foremost, I have not read many tutorials and such. I simply tried through trial and error what is working for me, so there is for sure a better, more efficient way and more knowledgeable people out there.

But I personally do things roughly like this:

1. Rip as many official images as possible (anime, artbooks, colored manga pages, etc.).
2. Check for additional images on imageboards and the web itself that are of high quality or beneficial for your character or design choice.
3. Sort your images and only keep the "best" ones, then upscale or rescale.
4. I usually try to get somewhere between 50 and 150 images with variations.
5. Set your keywords according to your model. Normally I do name and some core features like hair color, outfit, and hair style.
6. For training, I normally use Colab because I can let it run in the background of my laptop during work, and I kind of work most of the time q.q.
7. Finally, check the results, look at what went wrong or could be improved, and redo thingss until you're happy with the outcome.

A great lesson for me was that one time all my results had some kind of wet skin look by default, which is why I used too many images with "wet skin" without setting the correct keywords to separate those images.

I might write a real guide or make a video in the future if people are interested in this sort of thing.
But for now, I hope it helps you a bit.

I use "vae-ft-ema-560000-ema-pruned" and starting to include it after the initial release, because some people don't know about VAE and it is easier that way. So yeah, V2 comes baked with VAE.

@veryfirstquestion800 first of all, I am glad you like my model. About the VAE: Thats interesting, can you suggest me another one to try, because now I am really curious.

@zx96 @veryfirstquestion800
If I remember correctly... When I first started making V1, I tried: 560000, 84000 and the OrangeMix one, but somehow I liked 560000 the most and started including it because people were having problems installing a VAE themselves. And to this day, I still think it is the best option to include it in the checkpoint, because people get what they see in my previews right out of the box without additional tinkering/downloading.

Also at this point, I still really like the "bright & colorful" results.

And a personalized checkpoint like this is still based on personal preference at the end of the day, so yeah, it will most likely always stay that way, as I will also keep my preview "style" instead of pandering to a larger audience with a more normal anime-style, because that is how I am.

That said, I am at a different level of knowledge now, if that makes any sense, and I will be trying some different VAEs for future releases. Thanks for the reference link!

The best I can do for you guys is to see if I can find the core version without the VAE merge and upload it separately for you.

Sorry for the wall of text <3

I tested some different VAEs and might even found one that is better for Version 3 so thank you for that!

Also can you guys do me a favor? Please check the box in your stable diffusion settings --->

"Ignore selected VAE for stable diffusion checkpoints that have their own .vae.pt next to them"

Then choose the VAE you wanna use, and "apply settings".. The checkpoint should work now with the VAE you wanna use instead of the one included. I did some A/B testing and it did for me at least.

@veryfirstquestion800 thanks for explaining. Its still weird that it worked for me somehow.

But on a different note, I uploaded both versions on Huggingface so feel free to download from there: https://huggingface.co/KizukiAi/