Case Study: Super-Resolution and Photo Restoration

"Design the feature that makes a photo bigger, sharper and cleaner" sounds like the simplest prompt in this track — everyone has pinched-zoomed a blurry photo and wished it looked better. It is not simple, and it is a favourite senior interview prompt precisely because the same underlying problem — predict missing high-frequency detail from a low-quality input — ships as at least four genuinely different products with wildly different latency budgets: a phone gallery app upscaling a photo in under two seconds on-device (Google's Pixel Super Res Zoom, Adobe's Super Resolution), a cloud restoration tool spending tens of seconds to bring a scratched, faded family photo back to life, an e-commerce zoom feature sharpening a product thumbnail, and a real-time game renderer upscaling every frame in a few milliseconds (NVIDIA DLSS). A candidate who treats this as "pick one super-resolution model" has not understood the prompt. This case study is deliberately the block's edge-vs-cloud case: unlike the batch-offline product-photography case or the sub-five-second interactive cases (virtual try-on, generative fill), the central design fork here is not "cloud GPU pool, sized how" but "does this compute run on the user's phone or in our data center, and why."

This subject follows the same seven-step framework as case-study-fraud-detection and the other case studies in this track, with the generative-model vocabulary introduced at the altitude text-to-image-diffusion-models uses (state a loss and what it buys, skip the derivation) and every architecture choice argued against the GAN-vs-diffusion comparison table built in generative-adversarial-networks-and-face-generation. Aim to deliver the full answer in about 40 minutes, then use the follow-up questions to pressure-test it.