Description:

PWNy is just another PONY based model merge that started as a test and turned into a happy accident. Initially i wanted to remove some of the strong Waifu bias that many models carry and see if i can somehow steer the output more into a western direction. I don't know anymore how many iterations i went through, or how many different models and loras i tried to cook into it but some time in November last year it produced this model. And so far i'm quite happy with it and thought it is time to share it with you all, and also to have a backup online. 😁

Now, although it was pretty much what i was hoping to achieve it is also not without its flaws. Luckily for us the majority can be fixed with ControlNet or Lora's. So let's get to the Good and Bad.

The Bad:

• Strong style bias. This is mostly because it contains a mix of the F.Suarez & KryDecker loras by Sqweel & Kinky_Caterpillar.

• It sometimes refuses to output certain poses, but ControlNet can easily enforce your wishes.

• Somewhat bad at futuristic content, science fiction or cyberpunk for instance. Maybe i'm able to refine this in the future or someone else who knows more about this topic. Can to a certain degree circumvented by using more specific tags like 'cyberpunk edgerunners' instead of just 'cyberpunk'.

• Very rarely it can happen that the model tries to censor images. Can be fixed by using either or both rating_explicit or uncensored in prompts. Alternatively you can also use censored or rating_safe in the negative prompt.

• Last but not least, as always the bane of AI... Hands. Impact Pack has some stellar options to fix bad hands with segmentation and control net.

• On Site generation is subpar due to missing sampler options. Do yourself a favor and use the comfyui workflow provided in the previews, i tried to keep it as clean an easy to understand as possible.

The Good:

• It follows the majority of Danbooru tags well and doesn't require much input to output solid results. (See previews).

• Reacts to a lot of characters without the need for a lora (See previews), although a lora will give you much better fine control. Works with (photo)realistic lora's as well.

• Can do consistent characters quite well with no or minimal iterations.

• Can do environments but as mentioned above has problems with futuristic content.

• Supports a wide range of resolutions, both portrait and landscape. I usually use 1024 x 1280 for portrait and 1280 x 720 for landscape, but 1024 x 1536 works as well.

Technical:

Keep in mind, since i only use ComfyUI i can only help you with that.

• Steps: 10 - 30 »»» I usually use between 15 and 25 steps, above 30 makes little difference other than time.

• CFG Scale: 2.5 - 8 »»» I usually stick to 5.5 but lower or higher work as well.

• Sampler: DPM++ 2M SDE (GPU), DPM++ 3M SDE (GPU), Euler (A). »»» I usually use DPM++ 2M SDE GPU, but 3M and Euler should work as well.

• Scheduler: Align Your Steps, SGM Uniform.

• Resolutions: 1024 x 1024, 720 x 1280, 1024 x 1280, 1024 x 1536 and vice versa. But all common SDXL resolutions should work as expected.

• Prompting: At least score_9, in positive. 2d, 3d, realistic have a significant influence on the output in either the positive or negative.
  And only necessary tags when needed in negative.
  Don't limit your options by putting "source_pony, source_furry, source_anime," in the negative prompt by default. The tags used in the previews are meant as examples not as strict rules for optimal results, and also me being lazy.

Resources used:

Frankly i lost track in the process but here are some i recall.

• The OG of course - Pony Diffusion V6 XL

• The classic - AutismMix SDXL

• The highly underrated - Dr. Pibb PonyXL

• The chef cook himself - Vixon's Pony Styles - Detailed

!! PLEASE SUPPORT THE FOLLOWING ARTISTS DIRECTLY, if you can !!

• The amazing Fernanda Suarez - Fernanda Suarez Style by Sqweel

• The outstanding Krystopher Decker - Krystopher Decker Style by Kinky_Caterpillar

Legal Mumbo Jumbo:

I'd like to put this particular model under the 'Creative Commons Attribution-NonCommercial 4.0 International License'

• You are free to share, copy, redistribute, remix, transform, and build upon the model in any medium or format.

• You must give give appropriate credit to at least all the people mentioned under the Resources section above.

• You are not allowed to use the model for commercial purposes, this excludes Buzz on Civitai.

Please respect the last point of the license to show respect for the artists that have been used to train this model and all of its previous iterations.

Future Plans:

Honestly, i don't know.

Maybe i'm able to exorcize the issues this model has in the future, but i encourage anyone with more knowledge in that field to give it a spin.

And no, currently there is no Illustrious version planned.

Closing Words:

Feel free to review, post your creations or comment/ask questions.

And now go bonkers and have fun with it. 😊

P.S. This took me 8h to upload! Praised be the amazing internet infrastructure of Germany.