@eyesxz the problem is getting those 20 nice images from bangumi... most of them are badly cropped and a lot of good ones from the shows are actually missing

@eyesxz once I'm at the pc, I'll test a few charas that I'm suspicious about might not have enough knees and above shots.

@LittleJelly hmmm, have u ever seen what bangumibase is like? for most characters, finding 200+ good caps is even a easy thing... i dont understand what is you lots? have u seen 10k/100k/1m images from bangumibase? i have.

also plz dont expect to get good full-body images from screencaps, just find that in fanarts if u really need them.

@narugo1992 a lot of good shots are missing on bangumibase, its probably ok to use for school charas as there's plenty of shots, but on others i find it a bit lacking 🤷 also i spevialize on anime styled charas, and fanart sometimes brings a bit weird influence on the loras from my testing... But most importantly the charas are cropped really weird sometimes on bangumibase

@LittleJelly actually from my testing, for most charas with no less than 150 images in bangumibase, just train lora with it (even without any further cleaning) is 100% okay in all aspects.

but, for unpopular charas, it simply doesnt have so many shots. another low-key issue is that screencaps of unpopular charas often have much lower average quality than those of popular ones. most of the bangumi are like this, because it is not a good idea for the studio to cost tons of dollars on unpopular charas, not the problem of bangumibase. charas with 50- images in bangumibase usually cannot produce good lora from my testing.

so i suggest that for unpopular characters, do not use any form of screencaps as the main data source, because the face quality will be shitty. fanarts are much more reliable than screencaps for these charas.

btw, bangumibase's cleaning and further filering model is in my todo list, though the current data is more than enough for waifu lora training, but getting higher quality is always a win.