horror
RT @supercomposite@twitter.com
🧵: I discovered this woman, who I call Loab, in April. The AI reproduced her more easily than most celebrities. Her presence is persistent, and she haunts every image she touches. CW: Take a seat. This is a true horror story, and veers sharply macabre.
🐦🔗: https://twitter.com/supercomposite/status/1567162288087470081
horror
Genuinely horrified by this thread since to me "unreachable via text" and "very stable" (in both the woman's features and the gore theme) means there's a good chance that there's a lot of unlabeled images in the dataset depicting the abuse of one specific woman.
It's extremely common to (1) never really look at all your data—there's just too much!—and (2) to start with someone else's pre-trained model—it's too expensive to train from scratch!—with no way to know on what data it was trained.
