Every paint by number kit is the output of the same four-step problem, and understanding it explains almost everything about why your canvas looks the way it does. It also explains, precisely, why one photograph makes a good custom kit and another does not.
Step one: reduce the colors
A photograph contains millions of distinct colors. A kit has somewhere between twenty and forty pots. The first job is to decide which twenty-four colors best represent the whole image, and then to reassign every pixel to its nearest one.
This is called color quantization, and it works by treating each color as a point in three-dimensional space, with red, green and blue as the axes. Similar colors sit near each other. The algorithm groups those points into as many clusters as you have pots, and takes the center of each cluster as the color you will actually be given.
Two consequences follow immediately, and they are the source of most complaints about converted photographs:
- Smooth transitions become steps. A sky that graded imperceptibly from pale blue to deep blue now has four visible bands, because there are only four blues available.
- The extremes get pulled toward the middle. The very darkest shadow and the very brightest highlight are usually small parts of the image, so they end up inside a cluster whose center is less extreme than they were. The picture loses a little of its range.
That second point is worth remembering when you paint, because it is fixable. Adding a touch of the darkest color into the two or three deepest shadows, and the lightest into the two or three brightest points, restores what the math took out. It is the single biggest improvement available on a finished canvas, covered in how to make a paint by number look painted.
Step two: turn the colors into regions
Once every pixel has been assigned a color, neighboring pixels of the same color are grouped into contiguous shapes. Each of those becomes a numbered region.
The raw result of this is unusable. A photograph of a tree will produce tens of thousands of regions, most of them a few pixels across, because real foliage is noisy at the pixel level.
Step three: merge everything too small to paint
This is where a good kit is separated from a bad one, and it is a judgment rather than a calculation.
Every region below a minimum size is deleted and absorbed into whichever neighbor it borders most. The threshold is not set by what the printer can resolve; it is set by two physical facts about a person painting:
- The region has to be large enough to fill with a brush. Below roughly the width of a fine brush tip, it cannot be painted cleanly whatever the painter does.
- The region has to be large enough to hold its printed number legibly. This is usually the binding constraint, and it is why the numbers are the size they are. A region that cannot carry its own number needs the number placed outside it with a leader line, which is fiddly to read and fiddly to print.
Set the threshold too high and the picture turns to mush. Set it too low and you get a canvas that is technically correct and miserable to paint, covered in slivers you cannot fill and numbers you cannot read. Most of the difference between kits comes down to where somebody set this one number.
Step four: outline, number and print
The boundaries between regions are extracted as thin lines and printed onto the canvas along with a number placed inside each shape. The lines are deliberately faint, because they have to disappear under the paint. This is the reason pale colors need two coats: they are covering printed ink, and the lightest colors have the least covering power.
The trade-off that governs everything
Three things pull against each other, and you cannot improve one without giving up another.
Number of colors. Amount of detail. Size of canvas.
More colors means smoother transitions and more pots to keep track of. More detail means a closer resemblance to the original and smaller, more numerous regions. A bigger canvas relieves both, because the same regions get physically larger.
This is why we say so often that the larger canvas is the easier one. It is not a preference; it is the only variable that improves the other two at once.
Why some photographs convert and some do not
Now the practical part. A photograph converts well when it already resembles the output of the process above.
Works: one clear subject; even, directional light; distinct areas of color that a person could point at; strong separation between the subject and the background; a bit of contrast.
Does not work: flat overcast light, which gives the quantizer nothing to separate; several small faces, because facial modeling is continuous gradation with no boundaries at all, and boundaries are the only thing this process can represent; busy backgrounds, which eat the region budget; anything dark, where a dozen near-blacks collapse into two.
Faces are the honest hard case. A single large, well-lit face can be made to work. Four small faces in a group photograph cannot, at any canvas size, and no amount of care at the printing stage rescues it.
Our guide to choosing pictures for canvas painting goes through this with examples, and the custom kits themselves come in 12x16, 16x20 and 18x24.
Designed versus converted
One last distinction. A kit adapted automatically from an existing image inherits everything above. A kit drawn deliberately to be painted does not: the artist decides where the regions go, keeps them paintable, and arranges the colors so the palette separates cleanly on the table.
That is the difference you can feel in the first hour, and it is why the original 1950s kits, made long before any of this could be computed, are still pleasant to paint. Our piece on the history of paint by numbers covers where that came from, and our own designs are in the full collection.
