A block average or a centre sample, each named and measurably different, with the colour limit producing exactly that many colours and the count read back off the output.
Initializing in your browser…
Compare two images with a perceptual engine (pixelmatch) that ignores anti-aliasing, or an exact byte engine. Groups changes into regions with bounding boxes, reports SSIM, and handles size mismatches with an explicit policy.
Turn an image into a data URL whose payload is decoded again and compared byte for byte with the file, with the media type taken from the file's own signature rather than its extension and the size cost of embedding stated in full.
Whole-pixel crops with the rectangle typed or dragged, checked by decoding the output and comparing it with the pixels that went in. Quarter turns and mirrors are exact permutations, so nothing is clipped.
A one-pixel black and white checkerboard, pixelated at a tile size of 10. Its average is known before the tool runs, so the two ways of choosing a tile colour give two answers and only one of them can be an average.
Input
checkerboard.png 512 x 512 · tile 10 px
A flat field, at one of three values
centre pixel of the block byte 255 (solid white: the picture is gone) average, stored bytes byte 128 (the mean of black and white) average, light byte 188 (the byte carrying half the light)
Sampling the centre pixel of every tile is not an average and on a fine pattern it is not even close: every tile centre of this checkerboard is white, so the result is a blank page. The mean of the stored bytes is 127.5, and the mean of the LIGHT is byte 188, because byte 127 carries 0.212 of the light of byte 255 under IEC 61966-2-1. All three results are flat, so none of this is aliasing: they are three different questions, and the page names which one it answered.
Pixelating means replacing each block of an image with a single colour, and there are two different ways to choose that colour: the AVERAGE of the block, or one pixel out of it. They are not close. On a one-pixel black and white checkerboard the average of every tile is a flat mid grey and the centre pixel of every tile is pure white, so a centre-sampled pixelation of a checkerboard is a blank white page. Both are offered here and the page names which one it did, because the hard-edged centre sample is what pixel art wants and the average is what a mosaic or a privacy blur wants. The second choice is the space the average is taken in. Byte 127 carries 0.212 of the light of byte 255 under IEC 61966-2-1, so the mean of the stored numbers is not the mean of the light: on that same checkerboard the two answers are byte 128 and byte 188, which is 1.24 stops apart. Averaging the bytes is the default because it is what every other tool does, and the alternative is offered with the number beside it.
Select the rectangle over a face or a number plate; the rest of the photograph is untouched, and a large tile removes the detail rather than merely softening it.
The centre sample with a colour limit gives hard-edged tiles from a palette of a known size.
The block average with a circle or hexagon tile, where the gaps between tiles are part of the look.
Ask for 16 or 64 colours and read the count back: the output really holds that many.
A colour limit here means what it says. Asking for 16 colours builds a 16-entry palette by median cut (Heckbert, "Color image quantization for frame buffer display", SIGGRAPH 1982) over the colours the pixelated picture actually contains, and maps every pixel to its nearest entry by CIEDE2000, so the output holds at most 16 distinct colours. The page then decodes its own output and COUNTS them, and shows the count next to the number you asked for, because that is a claim a second program can check.
The usual alternative, snapping each channel to a step of 256 divided by the cube root of the colour count, cannot make that promise. Snapping for "16 colours" gives four levels per channel and so up to 64 combinations; snapping for "8" gives three levels and up to 27; and none of the resulting values is a colour anyone chose. The palette here is built from colours the picture contains and the count is measured rather than named.
Black and white is a choice of definition, not a single operation, and the page makes you pick one. A flat (R + G + B) / 3 puts pure red, pure green and pure blue all at 85, so at a threshold of 128 all three become black and a photograph of grass turns into a solid rectangle. Rec. 709 (ITU-R BT.709-6 item 3.2), which is the definition that belongs to sRGB, puts them at 54, 182 and 18, so green becomes white and the other two do not. Rec. 601 and CIE L* are offered alongside it with the document each comes from, and the threshold is adjustable.
A rectangular region can be dragged on the picture or typed as four numbers, so one face or one number plate is pixelated and the rest of the photograph is left exactly as it was: a pixel outside the rectangle comes out byte for byte identical. Four tile shapes are available; square is the only one that covers every pixel, and the other three leave gaps by construction, which is what makes them a halftone rather than a mosaic.
No. Every pixel in a tile is replaced by one colour and the rest are discarded, so there is nothing left to recover from the output file. That is what makes it useful for hiding something, and it is why a large tile is safer than a small one.
Yes. Drag a rectangle on the picture in the Region tab, or type its four numbers. Everything outside it is left byte for byte as it was, which the test suite checks against the source pixel by pixel.
The average is the mean of every pixel in the tile, which is what a reduction means. The centre pixel is one pixel out of the tile with the rest thrown away. On a fine pattern they give completely different answers: a one-pixel checkerboard averages to a flat mid grey and centre-samples to solid white.
Yes, and the page counts them for you. The palette is built by median cut over the colours in the picture and every pixel is mapped to the nearest entry, so the output has at most the number you asked for; the count next to it is the number the output actually holds, read back after encoding.
Because they give different pictures. A flat average of R, G and B puts pure red, green and blue all at 85, so all three fall on the same side of a 128 threshold. Rec. 709 puts them at 54, 182 and 18. Neither is a bug; one of them is the definition sRGB uses and the other is an accident, so the page makes it a choice.
Images are decoded, edited, and exported entirely inside this browser tab. No originals, exports, or metadata are uploaded.