The model imagines what it could look like before blurring. There is no information gain, it just makes a blurred image look nice using guesses.

Compare how a text to image model imagines what a whole image looks like based on a prompt and input noise.