Fix pixelated images with AI

What makes an image pixelated

An image looks pixelated when it holds too little information for the size you view it at. The plainest case is a small image stretched large with nearest-neighbor scaling: every stored pixel becomes a hard-edged square, and diagonal lines turn into staircases. Once a stretched copy like that is saved, the squares are baked into the file.

Heavy JPEG compression causes a different kind of blockiness. JPEG splits an image into 8x8 pixel blocks, discards fine detail inside each block, and often stores color at a lower resolution than brightness. Pushed too far, that leaves a grid of visible tiles, ripples called ringing around sharp edges and text, colors that bleed past their edges, and banding where a sky or wall should fade smoothly.

The damage also adds up. Every time a JPEG is edited and re-saved, or recompressed by a chat app or social platform, more detail is lost (this is called generation loss). A screenshot of a thumbnail keeps only the few pixels your screen showed, however large the original was. Zoom in close, and the pattern usually tells you which problem you have:

• Hard-edged squares across the whole picture: a small image that was blown up
• A faint grid of small tiles in skies and walls, plus ripples beside text: an over-compressed JPEG
• Soft, smeared edges with no squares at all: blur or low resolution, not pixelation
• A coarse mosaic over just a face, a plate, or some text: pixelation added on purpose

Why resizing and re-saving will not remove pixelation

A resize can only work with the pixels already in the file. Nearest-neighbor scaling copies each pixel into a bigger square, so the blocks simply grow. Bilinear and bicubic scaling blend neighboring pixels instead, which trades the squares for soft blur: smoother, but no more detailed.

Neither method can tell the picture from the damage, so JPEG tiles, halos, and baked-in squares are enlarged along with everything else. Sharpening afterward usually makes things worse, because it boosts contrast at every edge, including the edges of compression blocks. Converting to PNG does not help either: PNG is lossless, so it keeps the blocks exactly as they are.

How AI rebuilds a pixelated image

AI upscaling aims to undo the blocks rather than enlarge them. A model trained on a huge range of images redraws the picture: where your file shows a staircase, it draws the smooth edge most likely to have produced it; where a sky has broken into bands or tiles, it draws an even gradient; and where compression flattened a texture, it fills in one that fits.

Pixelperfect gives Google's image AI a short written instruction with each image. For pixelation, the key parts are "remove noise and JPEG compression artifacts" and "reconstruct crisp edges." Other parts pin the result to your picture: it is instructed to keep the subject, framing, composition, and colors as they are, with nothing added, removed, or restyled.

Two caveats apply. The detail is redrawn, not recovered, so lettering, logos, and faces only a few pixels wide can come back subtly changed. And the output has a set size: about one megapixel (roughly 1024 pixels on each side for a square image) when the original is up to about a third of a megapixel, and about four megapixels (roughly 2048 pixels on each side) when it is larger. So a pixelated image whose longest side is below about 1800 pixels comes back larger, while a big photo, which your browser shrinks before sending, can end up smaller than your original.

Upscaling cannot reveal what was pixelated on purpose

Faces, license plates, and documents are often pixelated to hide them. A censoring mosaic usually replaces everything inside each block with its average color, so countless different faces could leave the same mosaic behind. The fine detail inside each block has been thrown away, leaving an upscaler nothing to bring back.

An AI upscaler asked to depixelate such an image does what it always does: it draws the most plausible detail. The result can look sharp and convincing while being made up, so never use an "unpixelated" image to identify a person, read a plate, or reconstruct a hidden document.

Hiding something yourself? Use a fully opaque box, not a mosaic, blur, or see-through highlighter. A mosaic's block averages still carry a trace, so someone with a short list of candidates, such as possible plate numbers, can sometimes test which one fits; a solid fill leaves nothing to test.