A designer we interviewed wants to print classical and devotional paintings on shirts. The Clothing Design Studio already separated a painting into a few flat inks for screen print. This week we ran that separation on nine real public-domain paintings: Van Gogh's Starry Night, Munch's Scream, Turner's Shipwreck of the Minotaur, Hokusai's Great Wave, Matisse's Woman with a Hat, Picabia's Udnie, Sassoferrato's Madonna in Prayer, a framed Barocci and a small Munch Madonna.
Flat inks did badly. Even at eight inks, the average colour distance from the painting was 6 to 14 CIE76 delta E. About 2 is barely visible and 5 is plain to see. A painting is shading, and a flat ink has one tone.
Screen printers handle this with an index separation: the same few inks printed as a scatter of square dots, so the eye mixes neighbouring inks and the bare shirt into the in-between tones. The studio now has one. It uses the flat separation's inks, treats the garment colour as a free extra colour, and places one dot per finest dot the mesh can hold (0.3 mm by default) by error diffusion in linear light.
To compare the two, we blurred both the painting and the print the way an eye roughly blurs at a given distance (1 arcminute), then measured the distance. With four inks on a white shirt, the averages over the nine were:
| Seen from | Flat inks | Index dots |
|---|---|---|
| 0.5 m | 13.4 | 14.5 |
| 1 m | 12.3 | 10.6 |
| 3 m | 10.7 | 8.3 |
Up close the dots show, so index loses. From a metre it was closer on eight of the nine paintings. The one miss was a 184-pixel thumbnail, whose dots come out 1.5 mm wide. From across a room, four inks as dots (8.3) beat six flat inks (8.9). That is two fewer screens for the same look, and every screen costs money on every run.
Two things went wrong on the way. Plain error diffusion did worse than flat, because a painting's colours often lie outside what four inks can mix. The error then runs away into streaks. Clamping the corrected colour fixed that. Our second try, refitting the inks to what the dither asked of them, made things worse and was dropped.
The limits are plain. Neither method gets near "barely visible", because four inks cannot reach many of a painting's colours: dots remove the banding, not that limit. The eye model is crude. And no printer has pulled an index print from these films yet. Whether a single 0.3 mm dot holds on their mesh, and how much dot gain darkens it, is what a proof will show.
The tool is free and runs in the browser, and nothing is uploaded. The method and every number are in the venture's README under "Index separation for paintings", and eval/index-paintings.mjs reruns the measurement.