Super-resolution is the reconstruction of a detailed image from a lower-resolution version of it. It is a classically ill-posed inverse problem: for one small image there are infinitely many large ones that would look the same after downscaling, so the missing information cannot be derived, only filled in on the basis of learned knowledge about how real images look. That makes the method fundamentally different from interpolation, which merely computes intermediate values and necessarily blurs the result. Learned approaches began with the convolutional SRCNN, continued with residual architectures, and took a major leap in perceived quality with SRGAN, which replaced pixel-wise error minimisation with adversarial and perceptual losses; it turned out that optimising mean squared error averages all plausible solutions and thus yields a smooth, unnatural result. The newest systems build on diffusion models. It is essential to remember that the added detail is a probable guess, not fact – the model can invent sharp facial features or a legible licence plate that were never in the data, which makes the method useless in forensics.
Picture a restorer asked to turn a blurred photograph into a sharp one. He cannot “remove” the blur, because the information simply is not there. What he can do is draw on experience: he knows what eyes, brick walls and letters look like, so he paints very convincing detail into the blurred areas. The result looks great and is perfect for printing on a poster. But you have to admit what happened – the restorer partly made that face up. Which is why such an image must never be used as evidence in court, however much sharper it is than the original.