Color Palette Extractor

What 'dominant colours' means — and how quantisation finds them

An image with thousands of distinct colours doesn't have an obvious 'top 5' — you have to define one. Colour quantisation does that by partitioning every pixel into a small number of buckets of similar colour, then reporting each bucket's representative shade. This tool uses median-cut, the same classic algorithm behind PNG colour reduction and ImageMagick: it repeatedly splits the colour space along the channel with the greatest spread until it has as many buckets as you asked for, then returns each bucket's average. Crucially, median-cut is deterministic — drop the same photo twice and you get the same palette every time, which matters when a palette is going into a brand spec that has to be reproducible.

That determinism is also why this isn't k-means clustering. K-means produces comparable palettes but seeds randomly, so two runs on the same image can return slightly different colours — fine for art, frustrating for a documented brand palette. Median-cut trades a touch of theoretical optimality for repeatability, which is the right call here.

Dominant is not what your eye calls the 'main' colour

The single most common surprise: the largest bucket is often a colour you barely notice. A product shot on a white sweep is mostly white pixels, so white dominates; a screenshot is mostly its background chrome. The algorithm is counting pixels, faithfully — it just turns out the colour that covers the most area is frequently the backdrop, not the subject. That's why the 'ignore near-black / near-white' option exists: stripping heavy white backgrounds and dark UI before quantising stops the background from eating slots you wanted for the actual content.

The other limit worth stating plainly: this returns the colours that are in the image, not a colour scheme derived from them. It won't hand you complementary, triadic, or analogous palettes — for those, take the dominant hues and rotate them in HSL, or paste a base colour into a colour-theory tool. Extraction is observation, not generation.

Where extracted palettes actually earn their keep

• Branding from an asset: pull a defensible brand palette straight out of an existing logo or hero image when no official swatch sheet exists, then lock the hex values into the style guide.
• Data-viz and theming: derive chart colours or a UI accent set from a photograph so the interface feels of-a-piece with its imagery instead of bolted on.
• Design hand-off: export the palette as a PNG strip and drop it into a spec; pair it with the gradient generator to blend two dominant colours into a background, or feed values to the colour converter for print-ready CMYK.
• Mood and consistency checks: extract palettes from several candidate images to see which share a coherent colour story before you commit them to the same page.

Getting a clean palette: size and pre-processing

Match the palette size to the source. Four colours suit a tight brand palette where you want only the anchors; six is a good default for a hero image; eight to ten captures the range of a richly coloured photograph without descending into near-duplicate shades. Ask for too many buckets on a simple image and median-cut starts splitting hairs — returning three barely distinguishable blues that no human would treat as separate colours. The tool downsamples to about 200 px on the longest edge before analysing, which is deliberate: it's fast and, because dominant colours are an area statistic, the small loss of fine detail doesn't move the result. If a stubborn background still dominates, enable the near-black/near-white strip and re-extract.