Welcome!
Thanks for checking out my latest project! I spent a lot of time over the last week or two testing out possible updates for one of my other models. I ended up getting sidetracked, and "accidentally" created something that I thought turned out really well. As they say, beauty is in the eye of the beholder. If you try it out, please feel free to post your results and let me know what you think!
If you haven't used my merges before, my focus has been creating models that I really like the style of "out of the box", without needing to tweak the visuals with tons of LoRA or TIs. This model puts out some really nice looking 3d/realistic-ish images. It also carries a lot of the understanding of characters, creatures, etc baked into Pony.
All sample images were generated with no LoRA or TI. I use Adetailer and Hires.fix on all of my gallery images. Faces still generally look OK without Adetailer, but I recommend using it if possible.
If you are new to Pony, here are some tips:
Scoring - Use score_9, score_8_up, score_7_up, score_6_up, score_5_up in your positive prompt instead of SD1.5/SDXL quality tags. I have personally found that my images have much lower chance of frying if I keep the tokens below 200 (I usually shoot for under 150 when possible).
My base negative prompt looks like this:
score_1, score_2, score_3, score_4, score_5, score_6, watermark, text, signature, logo, monochrome, muscular, censored, ugly
Or use Zovya's Tis that do this for you with fewer keystrokes:
https://civarchive.com/models/332646/pony-pdxl-negative-embeddings
You can also use some of the baked in sources and content ratings to control the kind of output you get:
Sources - 'source_pony', 'source_furry', 'source_cartoon' and 'source_anime'
Ratings - 'rating_safe', 'rating_questionable' and 'rating_explicit'
My Usual Settings / Other Tested Options:
Sampler -DPM++ 2M SDE Karras / Restart / Euler A
Steps - 30--50 (30 for fast gens to test ideas, 50 for things I may save/post)
Resolution -896x1152, 832x1216 (or flip for landscape)
Hires.fix:
Upscaler - Ultramix_balanced, 4x_Ultrasharp
Steps - 15 or 20
Denoise -.3-.4
Upscale by - 1.5 - 2.0
Lastly, I'd like to thank the following amazing creators for their work on some of the components I used in this merge - Anzhc, insertusername, DucHaiten, and CitronLegacy!
Description
Refreshed from the ground up! Created using updated base model versions from v1.0, with some additional detail and art style tweaks on top. Overall seems to have better lighting, more detail, better hair/fur textures, better looking skin. Let me know what you think in the comments, and post some images to the gallery!
FAQ
Comments (4)
The new version is amazing, beautiful hands, beautiful details and works very well with suggestions and prompting :)
~ tyvm 4 ur work ~Looking for some tips... I use Serendipity to generate a lot of images. I like to setup a prompt with a lot of built-in dynamic-prompt options to get a lot of randomness then let it run all night. But one thing I struggle with is always getting the "default face". I've tried using [celebrity_name|celebrity_name] style of mixing faces, but it doesn't work very well. In fact, I was so disappointed in it, I decided to try just generating "close-up, portrait, (celebrity_name:1.5)" for my prompt (with the score stuff too) and then just running through my whole list of celebrities/models. I quickly realized that Serendipity/Pony doesn't "recognize" most of the names, and it doesn't do a reasonable job of imitating specific people... since that's not really my goal - I don't want those celebs specifically, I just want to mix up random attractive faces - anyway, I think you get the idea. So what I'm asking is how do YOU get "random" faces and avoid the "default face"?
Yeah Pony has lots of Anime/Comics/Cartoon characters that it understands, but not so much real people. The way I do it is by describing specific facial features, plus a random name and/or nationality. The amount of description I add usually depends somewhat on how close the face will be in the image, since I'm trying to keep the number of tokens from getting out of control. I created some really basic wildcards using ChatGPT a few months ago to help automate longer runs. I am finishing up a run of ten closeups with lots of description that I'll post in a minute.
@GhostBear1111 Thanks!
Details
Available On (3 platforms)
Same model published on other platforms. May have additional downloads or version variants.