A practitioner tech-report on training and evaluating style LoRAs as a sequence of measured decisions in a single feature space (DINOv2): calibrated per-dataset gates, controlled A/Bs with pre-declared adopt criteria, and metric design — with three illustration sub-styles (painterly, storybook-sketch, ink-wash) as the worked example.
Read it: https://cgr-ai.github.io/lora-methodology/
The site is a single page (index.md) rendered by GitHub Pages (Jekyll),
with all figures under assets/.