New attempt with the Lion + animefull-latest + LoHA algorithm combination, captions are also refined. Model stability has improved by much, yet detail level still feels far from ideal despite nearing overfit.
Alternative version trained with Lion optimizer is up. No longer overfitted and has better compatibility across different strains of checkpoint, albeit at the price of being somewhat less stable on details (thigh text, goggles). Forgot to review the caption files, so you still need to prompt each costume component to make it work.
Features KMS Blücher from Azur Lane in her High-Altitude Head Over Heels ski sport costume, trained on ACertainModel.
This LoRa started off as a humble trial-by-error side project on minimal dataset training, yet the outcome surpassed my initial expectation and more or less qualified itself for public release. Hope this gets you Blücher enjoyers out there.
Release v1.0: Applying within medium weight range (0.5-0.8) is recommended.
Release v2.0: Works well up to 1.0 full weight.
Release v3.0: Works well up to 1.0 full weight, image composition may benefit from a lower weight (give or take, 0.9) though.
Base: blcrazln, twintails, goggles, blue bodysuit, cropped jacket, back cutout, ass cutout, fur trim, two-tone gloves, sneakers
*Also try fake tail, clothes writing, crotch cutout if you'd like to, these are trained but have not been thoroughly tested.
FP16, dataset size 12 pics
Go ahead and upload yours!