Unpixelate an image — blocks back to detail

Shrinking an image to a few pixels keeps only coarse detail — the sampling theorem says finer detail is simply not stored ^[1]. Many different sharp images collapse into the same blocks, so an AI can only choose one plausible version, not recover the real one ^[2].

In 2020 PULSE, a research “depixelizer”, searched a face generator for any face that matched a pixelated input ^[3]. Given a pixelated photo of Barack Obama it produced a white man, and similar results for other people of colour — a vivid example of an AI inventing, not recovering ^[4].

It works when the image is simply small or badly compressed and you want a pleasant, larger version: Real-ESRGAN was trained on exactly that kind of damage — blur, noise, resizing and JPEG ^[5]. The less information is left, the more the result is the model's guess; creative engines guess the most ^[2].

Not by upscaling. Security researchers have shown that pixelated text can sometimes be guessed: Depix matches blocks against a reference image of the same font, size and colour ^[6], and Bishop Fox's Unredacter tries character guesses, pixelates each one and compares ^[7]. Researchers have also shown that neural networks can recognise faces through pixelation — picking among known candidates, not rebuilding the original ^[8].

The practical advice from those researchers is simple: to hide something, use a solid black box, never pixelation or blur ^[7]. An AI “unpixelate” gives a plausible picture, not the hidden original.