Note Feb 1, 2025:
V3 was trained using different trigger words and with a different set of captions. It produced better skintones, and far far lesser body-horror. Please note that the triggers now start with de51 (numbers in place of 's' and 'i')
Nov 18, 2024: V2 uses new trigger words since the old ones seemed to mix ethnicity with coffee flavors (yeah, go figure!).
Trigger words are different based on version
desilatte / desi_light (v2)/ de51 light (v3) - lighter skintones
desimocha / desi_medium (v2)/ de51 medium (v3) - medium skintones
desiespresso / desi_dark (v2)/ de51 dark (v3) - darker skintones
Concept segmentation is still patchy, though better with version 3. This only works sporadically at the moment and the dark skintones are not particularly dark. So this is more an experiment.
Would love to see someone test this out!
