Something worth separating when people talk about AI upscaling: restoration and invention are not the same operation, and most tools do both without telling you which one you got. Restoration recovers information that is still present in the source but degraded. Invention produces plausible detail that was never captured at all. For a holiday video nobody cares. For archival footage, security camera output, or anything that might eventually sit in front of a court, that difference is the whole ballgame. The uncomfortable part is that the metrics reward invention. Pixel error measures punish an honest blurry output, and the perceptual scores reward a sharp confident hallucination. So the optimisation target quietly pushes every model toward making things up.
BitFan
Public Service Atlas for Bittensor
