Pixel art color palette, read out of the picture

Because the file is not what your editor wrote, and three things cause it. JPEG is the commonest: it stores color lossily, so a flat area comes back as a dozen values that differ by one or two and every edge comes back as a gradient. The second is scaling — a sprite posted at 4x and saved back down, or any picture over 2,048 pixels on the long side, gets its neighboring pixels blended into colors nobody drew. The third is a soft brush or a feathered eraser, which leaves semi-transparent edge pixels that are counted at whatever RGB is stored under them. Export a PNG at native size from the original document and the count will match what you drew.

No — any pixel under 128 of 255 alpha is skipped entirely, so a sprite on transparency contributes only its visible pixels and the shares are computed against those. The threshold is a real edge case rather than a formality: a pixel at alpha 130 is counted at its stored RGB even though you can barely see it, which is why an anti-aliased outline adds entries and a hard one-pixel outline adds none. If you want the background counted, flatten it onto a color first and the flattened color becomes a normal entry.

A ramp is one material's shading sequence — the three to five shades that make up the skin, or the cloth, or the metal — and it is grouped by lightness because a well-drawn ramp does not hold its hue. Hue shifting is the technique that separates drawn pixel art from a lightness slider: shadows are swung toward blue and purple, highlights toward yellow and orange, so a single ramp can rotate thirty degrees or more between its darkest and lightest shade, and grouping by hue would cut exactly those ramps in half. What makes the chaining work is that a swerve in hue is charged three times what a climb in lightness is. Measured on real palettes, a genuine step between two shades of one material runs 0.15 to 0.22 in plain Oklab distance while a dark navy and a dark red sit 0.124 apart, so those two ranges overlap and no plain threshold can separate them — the a and b axes are compressed at low lightness, which makes every shadow look adjacent to every other shadow. Weighting the swerve pulls them apart, and the result was checked against palettes whose structure is already known: CGA comes back as each hue paired with its own bright variant, which is how the hardware defines it.

Yes, through either the .gpl file or the PNG. The .gpl is the GIMP palette format, which Aseprite, Krita and Inkscape also read, and it is written in ramp order with the Columns field set to the longest ramp so the entries land in something like their shading rows. The PNG is the other route, because those editors import a palette from an image by collecting every distinct color in it — which is exactly why the PNG here is one row of plain swatches with no gaps, no border and no caption. A black caption in the corner would arrive in your palette as a black entry.

Because median cut returns the pixel-weighted mean of each group it makes, and a mean of two colors an artist chose is usually a third color that is nowhere in the file. The reduction here already leaves the relaxation pass switched off, which stops the entries sliding further toward the center of their clusters, but that is not the same as guaranteeing they exist. The checkbox is: with it on, every derived entry moves to the nearest of the 20,000 commonest colors the file actually holds, and no two entries are allowed to take the same one. Turn it off when you want the representative color rather than a real one.

Between 8 and 32 for most work, and it is the ramps that decide it rather than the total. A character with skin, hair, cloth and one metal needs four ramps, and a ramp reads as a ramp at three shades and stops improving much past five — so four times four plus a couple of neutrals is 18, and that is why so many well-known palettes sit near 16. Jam limits are usually 4, 8, 16 or 32 for the same reason. A count far above that is not richer, it is usually a picture that was never quantized in the first place.

The deliverable is different: this page hands back the palette, and the reduce page hands back the picture. Both call the same median cut in Oklab from the same module, so a photograph reduced to 24 on either page yields the same 24 colors — but here they come out as a .gpl file, a hex list and a set of ramps with counts, and there they come out as a PNG of your picture painted in them. Use this one when the colors are what you are taking away; use that one when the image is.