/vt/ - Virtual Youtubers

>>60272557
Didn't prompt for messy hair when I made that image. That's exactly the technique i was talking about in that post. If I had removed the caption 'messy hair' from the training dataset and kept just 'long hair' as the only hair tag, her hair would have been almost always long, messy and grey when prompting for 'long hair' by tricking the text encoder as well that her hair should always be 'that'.
But I kept the tag so we can make her with other types of hair easier. If messier hair is needed, messier hair can be prompted. Unfortunately this first version is undertrained.