/vt/ - Virtual Youtubers

>>38603983
At this point Im starting to think its ALL about the quality of your dataset compared to quantity and variety. I cant help but think that more data is better than less data but higher quality data.

What even is high quality data when Im trying to pick images from Danbooru with only a few hundred to pick from, especially when they get cropped and the like and completely miss the tatoos anyway.

Apparently https://github.com/aria1th/Hypernetwork-MonkeyPatch-Extension there is an extension that allows for uncropped image use for training. It just limits whatever resolution is shortest to 512 pixels. But I havnt played around with it.

>spoiler
You're good mate, I enjoy throwing my thoughts at someone even if its probably more likely to be awnsered on /g/