Does anyone know why the text content in the images is weirdly off from the prompts? I couldn’t tell if that’s just a known limitation of the model, or a result of the author giving the prompts in some other language or something.
Dall-e knows how to draw things that look somewhat like writing, but doesn't actually know how to write.
It's because the model doesn't actually see text during training or prompting, it gets a compressed form called BPE instead, and the compression is fixed rather than learned.