原推：@SpookilyE Dall e generates images from text prompts, but if you ask for an image with text on it, the text will appear nonsensical.
It turns out that the text does imbed some “meaning” but only to the model.
It *is* correctly labeling the object, just not in a language we understand.
https://twitter.com/wintonARK/status/1531825867584077824