This LoRA is based on the 'double merge' technique outlined by the Reddit user shootthesound. Therefore the data was cropped twice, each at a different ratio/size, trained twice, and the selected LoRAs then merged at 100% each in Kohya.
Yes, I know this makes no sense, but read the post (archived here too), and believe - because this method obtains an order of magnitude better fidelity, and in particular leads to models where you can really ramp up the CFG without ruining the image quality.
This was my first experiment with the technique, and I have found that this character is the most disentangled I ever saw - it can convincingly perform poses that are utterly non-existent in any of the source data, and produce notably superior results at a wider range of steps and CFG settings.
HOWEVER - the way this technique works, you have to set the LoRA strength to 0.4 or less, or the results are trash. This is NOT (as it usually is) because the model was overcooked, but for reasons explained in the above-linked post).
See main post for further technical details about this LoRA.
