This is an experimental merge, aiming to create realistic images with the concept flexibility that Pony Diffusion V6 offers. May require significant prompt wrangling. YMMV.
This model is suitable for generaing both SFW and NSFW images.
Usage tips:
-recomemmended negative tag: makeup
For some reason the model puts heavy makeup on realistic faces if this tag is not given.
-use {character name} cosplay to generate pictures of people who look like the character instead of pictures of figurines or 3d render styled images
-specify the subject's age in the prompt, otherwise the model leans towards people who look around 30-50 years old.
Description
FAQ
Comments (5)
Any tips on prompting with this? Because this doesn't work like other photorealism checkpoints. Personally I have zero interest in ponies, furry stuff, cartoons, anime etc.
However I was intrigued by the dynamic poses/colours and foreshortening in your example images. I have found using the score tags works on this better than any negatives and helps with the disfigured face/cleft pallet look, mangled hands. Still trying out Lora's and textual inversions.
What I like about this checkpoint is that's is completely different from anyhting else (realism wise), it has very good prompt adherence and also does the amateur photo look well like talmendoxl.
Anyway thanks for this extremely interesting checkpoint. Looking forward to updates!
Here's my boilerplate for generating realistic images that I've built over the last year and half. I added the Pony Diffusion XL required positive prompt to the start.
Prompt: score_9, score_8_up, score_7_up,real, real life, sharp,crisp,clear, Nikon D800 BREAK
Negative prompt: makeup,doll,figurine,3dcg,lowres,low resolution,worst quality,low quality,normal quality,amateur drawing,sketch,watercolor,error,blurry,censored,jpeg artifacts,cropped,letterbox,letter box,out of frame,text,signature,watermark,artist name,username,monochrome,greyscale,black and white photo,bad anatomy,bad hands,missing fingers,extra digit,fewer digits,extra limbs,mutation,bad proportions,long torso
It's ok to save the negative prompt as a style, but positive prompt needs to be the first thing to work properly so having it in a style doesn't work.
I'm usually using DPM++ 2M SDE Karras or DPM++ 2S a Karras with 35 steps and CFG scale of 3-8.
Similar workflow to other stable diffusion models, generate a bunch of images at lower res first. Pick one you like and want to refine and tweak the prompt until you get something you like and then finish with hires fix.
Hope this helps.
@lostsnacks100 thanks for the reply. Since I posted, I found that ugly is a good negative for fixing deformities but makes the skin look plastic. Deformed-nose or disfigured-nose helps with the face. source_pony also helps with the face. And ya, makeup is a strong negative with this one. It's a bit like going back to 1.5 with all these negatives but it worth it as this a really interesting one.
I'll give your tips a whirl and hopefully they will help. Again, much appreciated!!
Great merge! Having a lot of fun with this.
Just scratching the surface, but it is great! 💖