💢 Cuteness is justice!! 💢
Hey everybody, it's been a while. I'm back with Sparkla this time, certainly a kindred spirit that understands that Maou-chan is the cutest in the world.
I'm testing a different approach to training this time (bunch of new settings, different tagging, etc etc). So expect a few more LoRAs in the near future and some updates to some of my older ones.
15/06/2026: Added Anima B1 version.
Give it a try! Have fun!
Recommended weight: 0.8 ~ 1.0
Default - Base Appearance
stell4sp4rkl3, purple hair, yellow eyes, hair over one eye, under-rim glasses, hair flaps, long hair, lantern hair ornament, purple star, large breasts,Default - Character Outfit
green beret, white shirt, three-quarter sleeves, chest harness, green capelet, green bowtie, leather belt, green skirt, medium skirt, side slit, brown pantyhose, leather gloves, brown mary janes, footwear bow,Additional tags / Accesories / Weapons
wheelchair, steampunk,There's literally a single angle for the Wheelchair in the Dataset (yes, the one from her game sprite). So, expect the wheelchair be super stiff and difficult to generate.
Description
Testing a different approach to training this time:
Ditched raw Kohya for LoraEasyTrainingScripts.
Changed network from LoRA to LoCON (Lycoris) with much lower ranks and max allowable batch size (around 50MB size for output files now).
Switch to CAME + REX annealing (no repeats, no warmup, but slight weight decay) for optimizer and scheduler.
Reduced datasets to <=30 images (only 15 images in this case) with full tagging.
So far, the results have been pretty good (have been testing them for the last month). I have like 4 different characters already trained with this setup now, posting them later this week hopefully.