Everyone got hyped when Stable Diffusion 3 went open source, but then when people actually tried it, generating humans was a total mess—limbs everywhere, twisted bodies, straight-up nightmare fuel everywhere. SD has always been the gold standard in open-source image generation, with a massive community ecosystem, so expectations for version 3 were sky-high, making the letdown even worse.
I make fan art with SD for a living, so messed-up character poses are basically a dealbreaker for me. Looking at community discussions, it seems related to the specific release version and how the training data was handled, but nobody’s really sure of the exact cause yet. It’s not completely unusable—landscapes and still lifes are supposedly fine—but for drawing people, we’ll have to wait for the community to fine-tune it for a while. Old SD users, don’t rush to fully switch over just yet.