Oh guess the AI folks haven't yet realized that you can increase perceived accuracy by adding a small amount of random noise.

Which of course they didn't, because then they'd need to understand some things about images instead of just throwing whatever data they can find at it and hoping it figures it all out from that.

Though I'm not sure how much it applies to images, as the examples I've seen were audio. But it's cool to hear a low bitrate audio sample of someone talking unintelligibly and then play that same clip with random noise also playing and suddenly you can understand what is being said.

The diffusion models at least were basically designed to "remove noise from a random image until a real image emerged" so that actually makes a lot of sense, interesting