Case study: a 20 × 20 high-contrast Einstein portrait

Turning six colours into a greyscale ramp

Ordered by perceived brightness, the cube palette is roughly white, yellow, orange, green, red, blue. That gives six usable steps of a grey ramp, and in practice a monochrome portrait only needs three or four of them: white for highlights, yellow for upper mid-tones, green or red for lower mid-tones, and blue for shadow.

This build used four values and deliberately skipped orange. Orange and yellow sit close enough in brightness that mixing both introduced noise in the hair without adding any structure, so removing one from the working set produced a cleaner result than using the full palette.

Why high-contrast sources win

The source photograph was lit hard from one side, which means the face already contained large regions of near-white and near-black rather than an even wash of mid-grey. Large flat regions survive quantisation; fine gradients do not. A softly lit studio portrait of the same person at the same size looked muddy in the preview because almost every pixel landed in the ambiguous middle of the ramp, where the quantiser flips between two colours on tiny brightness differences.

The practical rule this build confirmed: increase source contrast before uploading, not after. Raising contrast by about 20% and lifting the highlights slightly in any photo editor changed the mosaic preview more than any generator setting did.

Resolution and the moustache problem

The distinguishing features of this particular face are hair and moustache — both fine, high-frequency detail. At 16 × 16 the moustache merged into the mouth and the portrait became a generic old man. At 20 × 20 the moustache occupies its own band of stickers and the likeness snaps into place.

This is a general pattern worth planning around: the recognisability of a portrait is decided by its smallest identifying feature, not by the size of the face. Measure that feature as a fraction of the image width, multiply by the sticker count, and make sure it lands on at least three stickers.

Build data

400 cubes were bought in bulk with 20 spares. Sticker swapping averaged 85 seconds per cube once a rhythm was established, including peeling, placing and setting the cube in a tray — about ten hours in total, spread over five evenings.

The mosaic was built as four 10 × 10 panels. Square panels are easier to verify against a build sheet than long rows, because a miscount shows up as a visible break in the panel rather than a slow drift across the wall.

What we would change

The background was left uniformly blue, which is faithful to the source but makes the finished piece read as a floating head. Introducing a second shadow value — a scattering of red among the blue — would have added depth for no extra cost, since both colours were already on the ordered cubes.