Update: improved responsiveness to natural language prompts and conventional SDXL prompting style.

https://www.buymeacoffee.com/tripleheadg

Support me with a Pizza! Model making is hungry work!

I created a custom GPT4 for making easier SFW prompts! Check it out below!
https://chat.openai.com/g/g-ILX4kHo1I-pony-diffusion-guide


Pony requires a specific prompting syle, but I've jiggered it a bit so it doesn't really need it all the time, a mixture of natural language and tags should work much better now:

Rating_SFW/moderate/questionable/NSFW/Explicit

Add these to positive and negatives and it restricts it more generally to the types of content that you would like to see.

Also: "IRL" = realism.

And now you can combine score tags with specific concepts. For example: Score_9_eyes, Score_8_crowd, score_7_skin_texture, etc etc. Some things don't work, but when they do its a nice bonus!

You don't have to specify "Source_IRL/Source_anime" etc. just naming the style should be enough to push it in the right directions.

Usually I would add Rating_Explicit, questionable to the negative and SFW/Moderate to the positive at the same time. And this keeps things from getting to messy.

Also This model is very odd in terms of prompt weighting. You can use weights of up to 4.0 strength without making the images entirely awful.

So experiment with different weightings when it does not listen to you :D

--

Ponydiffusion was trained extremely thoroughly. Too thoroughly in fact. The lack of captionless training, or "caption dropout" is extremely evident and the consequences of such are far-reaching into the models use cases. Essentially limiting the model to only the longest and most complex prompting styles and/or raising the CFG levels to incomprehensible values to compensate.

I will attempt to rectify this issue in the model branch and make something effectively more useful than the existing main model. It won't be "better" per se. But it should be usable flexibly in a multitude of use cases.

Believe it or not I work incredibly quickly at merging models when I have feedback to work with.

This is a merge of PonyDiffusion, DreamState, Duc Haiten's Journey, AnimagineXL, Imperfect XL and Yamer's perfect design. Now merged with Autism Mix. Prony2.5D, Pyros NSFW 0.5 and ImperfectXL.

It does most of what PonyDiffusion does, but without the crazy ass "Pony Tags." It's more like an actual normal model, that does NSFW things, with ponies.

Some concepts are not working compared to PonyDiff, but this is a WIP and I will work towards getting most of them functional over time, without losing quality.