Single vs multi-image super-resolution

WHAT THIS TESTS The question checks understanding of where high-frequency detail comes from in super-resolution and how loss choice shapes perceived quality versus fidelity.

A GOOD ANSWER COVERS SISR versus multi-image: single-image SR has only one low-resolution input, so it must learn a prior and effectively hallucinate plausible detail it cannot truly recover. Multi-image SR uses several frames of the same scene with small sub-pixel shifts, so different samples of the underlying signal can be fused to reconstruct genuine high-frequency content, common in burst photography. Architecture: a modern SISR network stacks convolutional and residual blocks to extract features, then upsamples, often with a sub-pixel or pixel-shuffle layer, sometimes a GAN generator as in SRGAN. Loss contrast: L1 or L2 (MSE) minimizes per-pixel error and maximizes PSNR but averages over plausible textures, producing smooth, blurry output. Perceptual loss compares deep feature activations so results match high-level structure, and adversarial loss pushes outputs toward the natural-image manifold, both yielding sharper, more realistic detail at some cost to pixel-exact fidelity.

COMMON WRONG ANSWERS Saying MSE gives the sharpest, most realistic images, when it actually blurs. Or claiming SISR recovers true detail, ignoring that it hallucinates. Or treating perceptual and adversarial losses as identical.

LIKELY FOLLOW-UPS Why does MSE blur, the answer being it averages multiple valid solutions. What is the perception-distortion trade-off. How does pixel-shuffle upsampling work. Why does multi-image SR genuinely add information.

ONE CONCRETE EXAMPLE Upscaling one face photo 4x with an L2-trained network gives a smooth but soft result that scores high PSNR yet looks blurry, because the network outputs the average of all plausible high-resolution textures. Switching to an SRGAN with combined perceptual and adversarial loss produces crisp hair strands and skin texture that look far more realistic, since those losses reward outputs lying on the natural-image manifold rather than the pixel mean. The catch is the perception-distortion trade-off: some of the invented fine detail is plausible rather than faithful, so for tasks needing exact pixels a multi-image approach that fuses real sub-pixel information is preferable.