Compress PNG

PNG stores every pixel exactly, so it cannot discard detail the way JPEG does. What it can do is exploit repetition. Each row is first passed through a filter that encodes pixels as differences from their neighbours, and the result is then compressed with DEFLATE — the same algorithm as ZIP. Both stages reward predictability: the more often the same byte pattern recurs, the harder the file packs.

Reducing the number of distinct colours is what creates that predictability. A screenshot straight from a phone can contain forty thousand subtly different shades, most of them one-pixel variations invisible to anyone. Collapsing those into a palette of 256 carefully chosen colours changes almost nothing perceptually while making the pixel data dramatically more repetitive, and DEFLATE then does the rest. The container stays lossless; the image inside it is simplified first.

This is why PNG optimisation results vary so sharply by content. A logo with eight flat colours is already near the palette limit and may barely shrink. A gradient-heavy UI screenshot can lose 60% of its size with no visible change at all.

The right palette size depends entirely on how much continuous tone the image contains. Flat graphics survive brutal reduction because they never had many colours to begin with; anything with gradients, shadows or photographic content needs headroom or it will band — the smooth ramp breaks into visible steps.

Dithering is the mitigation. It scatters the rounding error across neighbouring pixels so the eye blends them back into a smooth gradient instead of seeing hard edges between palette entries. It costs a few percent of file size because it makes the data slightly less repetitive, which is precisely the trade: a little size back in exchange for gradients that hold together. Leave it on for anything photographic, turn it off for flat graphics where it only adds noise.

PNG is frequently used for images that should never have been PNG. A photograph saved as PNG is typically five to ten times larger than the same photograph as a quality-85 JPEG, and no amount of palette reduction closes that gap — the format is simply not built for continuous tone. If the image came out of a camera and has no transparency, converting it is a far bigger win than compressing it.

The reverse mistake is just as common. Screenshots containing text, logos with hard edges, and line art all look poor as JPEG at any reasonable quality, because the frequency-based transform smears sharp boundaries and leaves haloes around lettering. Those belong in PNG, where they are often smaller than the JPEG anyway as well as being pixel-exact.