Does anyone know if this is predicting the entire image at once, or if it's breaking it into constituent steps i.e. "draw text in this font at this location" and then composing it from those "tools"? It would be really interesting if they've solved the garbled text problem within the constraint of predicting the entire image at once.