Pixel Art Resizer

Every image resize has to decide what colour each new pixel should be. Bilinear and bicubic resampling — the defaults in almost every image editor — average the surrounding source pixels. That is exactly right for a photograph, where the underlying subject really is continuous and the pixels are just samples of it.

Pixel art is not a sample of anything. Each pixel was placed deliberately, and the palette is usually a handful of chosen colours. Averaging invents colours that were never in the palette and turns a one-pixel outline into a three-pixel gradient. Nearest neighbour instead copies the closest source pixel unchanged, so the output uses the same palette and the edges stay where the artist put them.

The tool keeps a smooth option for the case where you have a photo or a rendered sprite that genuinely benefits from it — but it is off by default, because for pixel art it is the wrong answer.

At 3× every source pixel becomes a 3×3 block. At 1.5× it cannot: 1.5 pixels is not a thing, so the resampler alternates — one source pixel becomes a 2×2 block, the next becomes 1×1. Nothing is blurred, yet the sprite looks subtly wrong, as if the art had been drawn on uneven graph paper.

The tool flags this the moment your target size is not a whole multiple of the source. If you need an in-between size — fitting a sprite to a fixed UI slot, say — scale to the nearest whole factor and centre the result, rather than accepting the uneven one.

A recurring problem with downloaded assets: the sprite is a 512×512 PNG, but the art inside it is 64×64 blown up 8×. Importing that into an engine wastes texture memory, and any further scaling compounds the block size.

To recover the original, the tool looks at the runs of identical pixels across every row and column. In a clean upscale, every run length is a multiple of the scale factor, so the greatest common divisor of all of them is the factor. A single stray pixel — an anti-aliased edge, a JPEG artefact, one manual edit — drops that divisor to 1, which is the honest answer: the image is not a clean upscale and dividing it down would destroy detail.

When it does find a factor, each output pixel takes the most common colour of its block rather than one sampled pixel. That matters when the image passed through a smoothing filter at some point: sampling a single pixel might land on an interpolated edge, while the majority vote cannot.

Input: PNG, WebP, GIF and JPEG — anything your browser can decode. PNG and WebP carry an alpha channel, so transparency survives the round trip. A JPEG source is already lossy: the compression artefacts around hard edges get scaled up along with the art.

Output: a lossless PNG at the exact target dimensions, with the alpha channel intact. The file name follows the source, with the new dimensions appended.