You've probably noticed most PonyXL realism models produce images with similar poses, lighting and an overly-pristine cleanliness that betrays the image was not taken by a physical camera.

It's common to find LoRAs around the concept of 'amateur photography' to add grain and skin textures to make the image appear less unreal. This can be a solution with some models, but still has the problem of not being able to pose the subject in a naturalistic way. Most images are not taken in a photo studio, they are taken in non-ideal lighting conditions and imperfect exposure/ISO settings.

StellaratorMix accomplishes the following:

Focus on realistic subjects in a variety of settings
Better understanding of physicality for arranging scenes/poses
Realistic anatomical micro details such as skin tension/folding and smaller muscles
Strong prompt adherence with or without LoRAs
Photorealistic materials, objects and backgrounds
Flexible and unique faces that do not regress into pony training at higher CFG
Defaults to high contrast, relatively dark images to solve the problem of low contrast bright images that tend to result from commonly used VAEs

Give it a try!

v2.0 Improvements:

Pruned size (11.3gb -> 6.4gb)
Greater concept awareness
Less frying at higher CFG
Improved prompt adherence
Tweaked contrast
Lighting (skin is more responsive to ambient light)
Better skin detail
More expressive faces
Better anatomy
Better clothing detail

You can still download and use the 11gb v1.0 model if you prefer the slightly darker images that it produces. Many of the more subtle changes can still be achieved through prompting, but v2.0 has generally better out-of-the-box performance.

If you are having issues with the built-in VAE use fp16-fixed instead.