Anyone else noticed the Venus poster having actually legible text? That’s very cool. Previously I’ve only ever seen generated images having garbled text.
I asked DALL-E 2 for several pictures of "DO NOT ENTER" signs and a few of them had signs that said "DO ENTER".
Almost! The subtext is still in a dreamstate
I'm surprised Stable Diffusion and others haven't cracked legible text by now. OCR has been solved for decades now, surely they could include some OCR term in the loss function while training it?