What this color palette generator does differently
Most palette extractors count pixels. They tally the exact RGB values in the image, sort by frequency, and return the top few. On a real photograph that produces a palette of five nearly identical shades, because the most common exact values are all micro-variations of whatever fills the largest area.
This one uses median cut. It treats the image's colours as a volume in three-dimensional space, then repeatedly splits the widest-spanning box in half along whichever channel varies most, until it has as many boxes as you asked for. Each box's average becomes a swatch. The result spans the image rather than clustering in one corner of it, so a sunset gives you the sky, the water and the silhouette instead of three slightly different oranges.
It is also deterministic. The same image always produces the same palette, which matters when you are matching an existing design rather than exploring.
Reading the percentages
Each swatch shows the share of sampled pixels that fell into its group, and that number carries information the colour alone does not.
A colour at 40 percent is the frame's foundation, the thing a viewer registers as "what colour is this image". A colour at 3 percent is an accent, and accents are usually the interesting part: the one warm light in a cold scene, the single saturated prop against a muted set. When you are building a scheme, the large percentages give you backgrounds and the small ones give you the highlight that makes the scheme work.
Building an image color palette from a film still
A color palette from image is the most common thing people want here rather than a scheme generated from nothing, and this case deserves its own note.
A still from a film you like gives you a colour scheme that has already been proven to work on screen, complete with the proportions the colourist chose. Grab six colours, keep the percentages in mind, and you have both a palette and a weighting. Feeding those hex codes back into a prompt as explicit colour instructions is far more reliable than describing a mood and hoping for the same result.
Sampling and privacy
The image is downscaled to a 128 pixel grid before its pixels are read. A 24 megapixel photo would be 96 MB of raw colour data, and quantizing that would freeze the tab, while the palette from a small thumbnail is visually identical. All of it happens in this browser tab, so the image is never uploaded.