Fine Tuned on cherry picked images from PersephoneHentai and PersephoneComicArt.

I'll still be working on improving this as i go.

None of the preview images are cherry picked.

If anyone has any spare buzz id appreciate it, completely tapped out from training.

-V3-

------

Same as v2 this improves the hands quite a bit, along with small details and such. It was refined using more realistic images than previous so its 3d capabilities have improved as well. its still capable of essentially recreating v2 images with minimal changes, just more details. At least in my testing.

Since the main focus is 2d, for 3d you have to prompt quite heavily for it. It can do it with quite a bit of detail but its not the main focus of the model.

Again look at the preview images for prompt help if you want.

I'm just majority sticking with Euler a now but 2M SDE is still an option if you want more detail, the settings haven't changed.

I'd honestly say its more like a version 2.5 but I don't really like the half numbering system so its just v3.

For the next version im gunna try get the model back to more 2d, It kind of had to be refined on 3d/realistic stuff purely for the details and to make the hands and stuff not terrible but that also pushed it away from 2d. Since everything seems to be in a good spot im gunna try get the 2d style back for the next version.

-V2-

------

This version has an improvement on the hands, prompt adherence and in my opinion overall image quality.

It did lose some of its 2d anime aesthetic so just prompting anime style, 2d in positive and 3d, realistic in negative should bring that back for the most part. In my opinion the slight loss in style was worth the improvements.

Recommended Settings

Euler A, for some reason this model doesn't really like DPM based schedulers, they work but 2M SDE causes censoring. If you don't care about nsfw and having to prompt away the censoring then 2M SDE is better overall quality.

If you do decide to use 2M SDE then putting censor bars in negative should negate the censoring.

CFG: 7-9. personally I'm using 8.

SDXL VAE is baked in so theres no need to use one.

UnaestheticXL_Alb2 causes the images to be cooked for some reason. All other negative embeddings are fine just avoid that one, unsure as to why its only that one it happens with.

All the other settings remain the same as listed below.

I added some images in the previews using loras just to show that its fine with loras, plus everyone seems to be using that one lora so figured I may as well show it works with this model too.

-----------

Recommended Settings

I still suck at writing descriptions so this will just be general settings.

Use regular SDXL vae

Regular SDXL resolutions

Euler A, DPM++ 2M,

CFG 5-7

30-55 Steps

I use Adetailer for faces unless the face is close, when the face is close to the screen Adetailer is usually a quality hit.

Recommend Hires Fix for higher quality images but thats just standard now. 0.35-0.4 denoise and 10-15 steps. I use 4x_NMKD-YandereNeoXL_200k upscaler

Recommended to add Anime Style after the score prompts if you wish for a more well...anime style.

May have to write Censored in negative if you start to get censored imaged. i didnt fine tune with any censored images but it censors sometimes.

2M SDE can cause some weirdness, mainly censoring.

Since its a pony model the score prompts are pretty much required. Just prompt it like a regular pony model.

V1

Not entirely happy with the hands. Most of the time they're good but sometimes they're a bit weird. Not consistent enough for my liking. I'll be working on improving that. In the meantime you can use negative embeddings to reduce the frequency if you wish.

Alternatively use Good Hands for Pony on about 0.7