Belko Style [PonyXL] [8MB]

I couldn't believe there was no Belko LoRA for PonyXL, so I made one myself. Definitely worth the effort.

Style LoRA to replicate Belkos art style, based on PonyXL v6, though I recommend using it with AutismMix, for best image quality.

Since Belko is mainly an NSFW artist, the model does best with NSFW images, but is also able to generate SFW images and everything in between. It is somewhat biased towards paizuri content (because that was Belko does best), but you can usually stop it by simply adding "paizuri" to the negative prompt.

If you like to the LoRA, please leave a reaction, review and/or tip. Enjoy!

Usage

• No activation/trigger needed. Simply include the LoRA in the prompt.

• Recommended to use at strength 0.8 - 1.0

• Use Clip skip=1 or Clip skip=2 (other values may also work, use at own risk).

• Works best at 1024x1024. Struggles to get the style right at lower resolutions.

• Recommend to use with SDXL VAE or flatpiecexlVAE (depending on image content and preferences)

• Optional: Using "source_anime" can help the style, but is not required.

Example

See preview images for lite version.

Unpruned vs. Pruned vs. Lite

I offer three variations of the model, which only differ in the file size and quality. Pick the one that best suits your preferences. If you can not decide, simply pick the "pruned" version. Note that the quality is relative, even the lite version usually produces images of good quality.

• unpruned: The originally trained model. Largest file size, best quality. Pick if you want the best possible quality, at the cost of disk space.

• pruned ("normal" version): Downsized model. Almost same quality as the unpruned version, with noticably smaller file size.

• lite: Drastically downsized model. Smallest file size. Still has good quality, but noticeably worse than unpruned and pruned

Notes

This LoRA was more effort than I expected. Removing the watermarks was a lot harder, since Belko likes to tag "inside" the image instead of at the edges. Even though I removed all watermarks I could find, some must have slipped through, because the LoRA will sometimes create watermarks, though not significantly more often than the base model.

In an attempt to reduce the paizuri bias, I split the training data and reduced the "weight" of the paizuri images. I think it worked? I can't really tell, but I'm very happy with the results either way.

I also tried to use 10k steps instead of 5k and compare the results. The 5k version is "softer" while the 10k version is more stylized. For now, I will only publish the 10k version. Maybe I'll revisit the 5k version at some point.

I also expected the model to be absolutely horrendous at creating hogs, considering that they are always censored in the training data and I was too lazy to uncensor them. But surprisingly, the model is still pretty decent at generating sausages.

Using training data

Feel free to contact me for access to the training data.
The training data (images and tags) has been sourced from rule34.xxx and cleaned up manually by me. When using the training data or modified versions of it to create models or datasets, please credit both me and the original artist.

Disclaimer

This LoRA is not affiliated with the artist it is based on. If you like the art style, checkout the artist on their socials:
Pixiv: https://www.pixiv.net/en/users/39123643
Twitter: https://twitter.com/ZURIFFIN