Zero IQ style attempt with a tiny dataset on an EZ-Bake lora trainer.

Current issues the closer you get to weight 0.9 and up:

• Test.1 - Apparent bias towards long nasal lines due to how it's interpreting the original material.

• Only, they're full on straight vertical lines, dead center of the face.

• Test.2 - Super low nose placement.

• Both - Probably handles other areas (like midriffs & navels) weirdly also, given one particular image in the set.

Went back with manual adjustments (and a bit of padding) to Test.1's set for a version "1.5" after results from 2. Minus the image providing the midriff issue, which was not taken into account at the time and should be present in all versions.

Plan to combine the adjusted Test.1 set to Test.2 for a final 2.5 to see how that turns out, maybe go back to 1 again to change that one pic.