Free Paint by Number Generator

For a line drawing without the colour regions the coloring page generator is the tool, and to copy a picture by hand instead the rectangular grid tool divides it into cells. The commercial form was created in 1950 by Dan Robbins, working for the Palmer Paint Company in Detroit, who has said the idea came from an account of Leonardo giving numbered sections to apprentices to complete. The kits were an immediate commercial success and became a cultural fixture of the decade.

Quantisation means finding a small set of colours that represents a much larger set with the least error. k-means does this by iterating: assign every pixel to its nearest centre, move each centre to the average of the pixels assigned to it, repeat until it settles.

The weakness is the starting position. Seed the centres at random and two of them frequently land close together in a photograph's dominant tone — a portrait is mostly skin and background, so random seeding gives several skin bins and none for the eyes. The palette then has twelve entries and eight useful ones.

Seeding by farthest point fixes this. The first centre is taken from the image, and each subsequent centre is placed at the pixel furthest in colour from every centre so far. The result spreads across the picture's actual colour range before the iteration begins, which is why small palettes still capture the dark accents that carry a photograph.

The clustering here runs in RGB, which is not perceptually uniform — equal distances in RGB do not look equally different. Working in a perceptual space such as CIELAB gives measurably better palettes at the same colour count. RGB is used because it is fast enough to re-run on every slider move, and the difference matters less once the count is above about ten.

Straight quantisation produces a label for every pixel, and along any soft gradient the labels alternate. A cheek shading from one skin tone to the next becomes a band of interleaved pixels rather than a boundary, and outlining that produces a field of noise.

A majority filter replaces each pixel's label with whichever label is commonest in its 3×3 neighbourhood. Isolated pixels are outvoted and absorbed; a real edge, where one side genuinely dominates, survives. It is a crude operation and it is the single biggest improvement to the output.

This is the distinction most generators miss. Twelve colours does not mean twelve areas — a single blue might appear in the sky, in a reflection and in a shadow, as three separate places on the sheet. Numbering by colour would put one "3" somewhere in the picture and leave the other two areas unmarked.

Finding connected regions requires a flood fill: start at an unvisited pixel, spread to every neighbour sharing its label, and record the extent. Each region found that way gets its own number at its own centre of area, which is what makes the sheet followable.

The centroid is not always inside the region — a crescent's centre of area falls outside it — but for the compact regions that survive an area threshold it is reliable, and it is far better than placing the number at a bounding-box centre.

A photograph quantised to twelve colours produces thousands of connected regions, most of them a handful of pixels along a boundary. They are real in the data and meaningless on paper: nobody paints a four-pixel island, and a number printed on one is illegible.

Setting a minimum area and leaving smaller regions unnumbered is what converts the output into a template. It loses fidelity, deliberately. The right value depends on the printed size — a sheet printed at A3 can carry smaller regions than the same template at A5 — which is why it is a control rather than a constant.

The subjects that work have strong, simple forms with clear tonal separation: portraits with directional light, animals against plain backgrounds, buildings, boats, still lifes. Anything where you could describe the picture in a dozen shapes will quantise into a dozen paintable areas.

Texture is the enemy. Foliage, fur, gravel, crowds and water surfaces all carry their information in high-frequency detail, and quantisation turns that into thousands of tiny regions. Raising the minimum area helps but eventually removes the subject along with the noise.

Low-contrast photographs are the other difficult case. If the whole picture sits in a narrow tonal band, the clustering has little to separate, and the template comes out as a set of shapes that look almost identical when painted — technically correct and visually flat.

The plain drawing grids cover the by-hand alternative, and the square grid page is where that starts. The palette list gives each number its colour as a hex value, which is a starting point for mixing rather than a paint-shop code. Matching acrylics or oils to a hex value is a judgement, and it is usually easier to mix toward the finished-look preview than to match the swatch exactly.

Working from the largest regions to the smallest, and from darkest to lightest, is the conventional order. It keeps the sheet readable for longest, because the numbers on the small regions are the last thing you need and the first thing to be covered.