Teaches AI how to do underwater stuff.
The LoRa has mainly learned artist styles "@amanchu!", "@artbysaver" and "@rou22". But the model is flexible enough that it can work with other styles too. Just prompt for some style.
Scuba gear is a bit too complex perhaps and can often look bad. If it gets in the way negative prompt helps. Scuba gear was mainly captioned as "scuba", "diving regulator", "scuba tank", "flippers", "diving mask", "snorkel", "diving belt".
Older Illustrious version available at: https://civarchive.com/models/1317934/underwater-style-mix
Description
Trained on Anima preview 2.
I tested doing training a bit different for this. First I threw all 2556 images, without considering quality or balance, at it for 3000 steps (4 images per step), then resumed training for another 3000 steps with curated 820 image set, then resumed once more for another 2000 steps with same set. Results seem pretty good.