Weird model tuning merge thing.

It's kind of hard to explain exactly how, but through some incorrect merge coding, the initial version of the stack was able to "fine tune" models rather than straight up creating new models (though you could achieve that if you went far enough.).

Through this experimentation I accidentally, instead of merging the models as expected, fine tuned Niji SE and improved the anatomy. I'll show a couple of stages of the transformation below for you all to look at and then some comparisons between Niji SE and what I call "Niji Unbaked."

Anyone who wants to know more about the methods behind it, just hit me up.


Besides from the cover image, the first 5 images are the tuning process, I generated a 512x512 image at 2.5 CFG purposefully found a seed which produced obvious artifacts and then accidentally found I was cleaning those artifacts out of the image with further settings changes.

Once I realized what was happening I decided to work on refining this further, without deviating too far from the original images composition. And eventually it came to an image result which seemed both similar enough to the original output of NIJI SE and clean and decent enough that I wanted to save that as a full checkpoint. Upon further comparison at the correct CFG's for the Niji model with the recommended settings I found that this new version maintained those improvements across the board.

The images after the first 6 will all be:

Niji SE, Followed by a comparison to my custom Niji unbaked version.