Upscaling

Traditional upscaling stretches an image and interpolates between the pixels it has, which gets you a bigger, softer picture. AI upscalers go further and invent plausible detail from a model trained on sharper footage. Both are useful for old material, phone footage, or archive clips that have to sit next to modern shots. Neither recovers anything. Detail that was never recorded is being guessed at. On faces and text the guesses are the first thing a viewer notices, and grain, compression artefacts and motion blur get amplified along with everything else, so a soft, noisy clip becomes a large, soft, noisy clip.