How Image Upscaling Really Works: From Bicubic Interpolation to Real-ESRGAN to Generative Super-Resolution

Take a small, blurry photo and try to make it bigger. For most of digital imaging history, “bigger” and “actually sharper” were two different things, and no amount of resizing could turn one into the other. That’s changed — but it’s worth understanding exactly what changed, because “AI upscaling” spans a few genuinely different techniques with different tradeoffs, not one magic button.

The old way: interpolation can’t invent what isn’t there

Classic upscaling methods — nearest-neighbor, bilinear, and bicubic interpolation — all work on the same basic principle: given the pixels you already have, mathematically estimate what value new, in-between pixels should take. Nearest-neighbor just duplicates the closest existing pixel, which is why it produces obviously blocky results. Bilinear and bicubic are smarter, blending values from a small neighborhood of surrounding pixels using weighted averages, which produces smoother, less jagged results.

But “smoother” is the operative limitation. These methods have no information about what the image is actually a picture of — they don’t know a blurry patch of pixels is a human eye or a strand of hair. They’re purely mathematical estimators working on numbers in a grid. The result of interpolating a low-resolution image up to a larger size is a bigger image with the exact same amount of real information — just spread out and smoothed, often looking soft or slightly blurry rather than sharp. You can make an image larger this way. You cannot make it sharper, because sharpness requires detail these methods have no way to generate.

The deep-learning shift: models that learn what detail should look like

Real-ESRGAN and similar deep-learning super-resolution models work on a completely different principle. During training, the model is shown enormous numbers of image pairs: a high-resolution original, and a deliberately downscaled (or otherwise degraded) low-resolution version of the same image. The model’s job is to learn a function that maps the low-resolution version back toward something close to the high-resolution original — and crucially, it learns this across huge, varied datasets, so it isn’t memorizing specific images, it’s learning general patterns: what edges typically look like at high resolution, how fabric weave, skin pores, foliage, and hair typically resolve into fine detail.

At inference time, when you feed the trained model a new low-resolution photo it’s never seen, it applies those learned patterns to synthesize plausible fine detail consistent with the low-resolution input — essentially a highly informed hallucination of what the high-resolution version probably looked like, grounded in everything the model learned from millions of similar training examples. This is why AI upscaling can turn a blocky, blurry face into one with convincing, sharp-looking skin texture and defined hair strands: that detail is being generated based on learned patterns, not recovered from data that was never there in the low-res file.

The newer frontier: generative upscaling

The most recent wave of upscalers uses generative image models — architectures related to the diffusion models behind AI art generators — rather than purely discriminative super-resolution networks like Real-ESRGAN. These models are even more capable at inventing convincing fine texture, because they’re drawing on the same kind of rich, generative image priors that let diffusion models produce entire photorealistic images from scratch.

The honest tradeoff here is worth stating plainly: the more generative and creative an upscaler is, the more it’s capable of inventing detail that looks fantastic but wasn’t strictly implied by the original low-resolution input. A generative upscaler might render a slightly ambiguous blurry patch as confident, sharp fabric texture — and it might guess wrong about what that texture actually was. This is fine, even desirable, for enlarging a casual photo where “looks great” matters more than pixel-perfect fidelity. It’s a real consideration for anything where faithfulness to the actual original content matters, like forensic or documentary use, where a purely reconstructive (rather than purely generative) approach is more appropriate.

Three approaches, three tradeoffs

Lined up together: classic interpolation is fast and perfectly faithful to the original data, but can’t add detail — it just smooths what’s there. Discriminative super-resolution models like Real-ESRGAN add genuinely new, learned detail while staying closely anchored to the low-resolution input’s actual structure. Generative upscalers push furthest into synthesizing convincing fine detail, at the cost of sometimes drifting further from strict faithfulness to the source.

Where this fits

This is exactly the territory ClearUp AI operates in — free upscaling to crisp 2K or 4K, built on the deep-learning super-resolution approach that reconstructs plausible fine detail rather than just stretching and blurring pixels the old-fashioned way.