Why Upscaling Cannot Invent Detail

Upscaling turns a smaller picture into a larger one, but the information available never increases. What differs between methods is how convincingly they conceal the shortfall.

Missing samples cannot be reconstructed

A picture recorded at one resolution contains a fixed number of samples of the scene. Detail finer than the sampling interval was never captured.

Enlarging asks for values at positions between those samples. Any answer is an estimate, because no measurement exists for those positions.

This is a hard limit rather than a software shortcoming. A better algorithm produces a better-looking estimate, not a recovered measurement.

Conventional methods trade softness against artifacts

Simple averaging produces a smooth but blurred enlargement. It is predictable and never introduces anything false.

Sharper methods weight nearby pixels to preserve edges, which looks crisper but can create faint halos beside high-contrast boundaries.

Neither invents texture. An enlarged picture of a brick wall shows the same bricks, larger and softer, with no additional mortar detail.

Learned methods synthesize plausible texture

Models trained on many images can generate detail statistically consistent with what a real scene at that resolution would contain. The output can look strikingly sharp.

What they produce is plausible rather than accurate. The detail is invented in the sense that it does not correspond to what was in front of the camera.

For entertainment this is usually acceptable and often preferred. For anything where the picture is evidence, invented detail is a serious problem.

Source flaws are enlarged too

Compression artifacts, noise and sharpening halos all scale up along with the picture. A method that sharpens will sharpen those as well.

Learned methods can misinterpret artifacts as scene content and elaborate them into confident nonsense. Blocking edges become architectural features.

Cleaning the source before enlarging usually matters more than the choice of method. Enlarging a damaged picture magnifies the damage.

Downscaling behaves entirely differently

Reducing resolution discards samples, and doing it properly means averaging groups of them rather than picking one. Averaging preserves an accurate impression of fine detail.

Picking rather than averaging causes aliasing, where fine patterns turn into coarse false ones. Striped fabric shifting into broad moving bands is the familiar symptom.

Correct downscaling of a high-resolution source usually looks better than native capture at the lower resolution, because averaging suppresses noise. Shooting larger and delivering smaller exploits this.