After my first release, I quickly realized how impractical it was to share my style with a LoRA that is very restrictive as to which base model it works on. So, thanks to everyone's support in my first week here on Civitai, I am proud to present a generalized Pony LoRA for my style!

With this release, you can now use my style directly on Civitai with the base Pony model and some of your favorite Pony derivatives!

Here are the settings I have found to work well, on Pony Diffusion V6:

• CFG: 3

• Steps: 20-30

• Sampler: DPM++ 2M SDE Karras or Euler a

• LoRA Strength: 1

I am so happy with how well Pony was able to learn my style with this LoRA. It works well using my typical simple prompting style. I am hoping that more of Pony's strengths can be incorporated into use of this LoRA to maximize its quality.

Please note! If you want to achieve results close to my examples, do not use Pony tags such as score_9, source_anime, etc.! Using those tags can have interesting effects, but can also dilute the style significantly!

V2 Update! Available now:

• Excessive monochrome outputs have been fixed!

• The style is much cleaner now, and outputs are closer to my original ConRad XL + LoRA. This works with my simple prompting style even better than in V1.

• V2 showcase exclusively uses simple prompts - no Pony score or source tags.

Challenges:

Pony seems to really want to give catgirls human ears in addition to cat ears! Please let me know if you are aware of a way to counteract this!

Compared to the specialized LoRA, it feels a bit weaker at generating eyes that aren't close to the camera, but it still inpaints very well. Adding "quality" tags, or Pony score tags, does seem to help with this.

V1 seems to give monochrome outputs more often than it should, and in some seeds it can't even be countered with negative prompting. Removing monochrome images from the dataset fixed this in V2. Including some color details in the positive prompt usually avoids this too.

I am woefully inexperienced with Pony as a base model, so my showcase features my simple prompting style. With V1, I started to experiment with Pony-specific scoring and source tags. Using them can drastically alter the art style, but it does seem like there could be a nice balance to be found. Adding "source_anime, source_cartoon, score_9" to the prompt seems like a good starting point. Adding "realistic, 3d" to the negative prompt generally seems to lessen the effect of Pony tags on the art style. In V2, going without Pony-specific tags results in a style very close to my original goal!

-

This model was trained on outputs from my previous model set: ConRad XL and the original WildSwanku LoRA.

-

Please enjoy, and I'd love to see what you create with this LoRA! I'd also like to see the results of experiments on various Pony-based models. Show me your favorites!