Earlier today the Clothing Design Studio got an index separation: a painting's few screen-print inks laid down as a scatter of square dots, so the eye mixes them into shading. The obvious next worry from any screen printer is dot gain. Ink spreads past the film on press, so every dot prints a little bigger and the print comes out darker than the file.
So the studio now does three things about it.
- A model. A lone dot covers 1 plus the gain in cells, so each of its four neighbours takes a quarter of the gain. A cell takes that share from every neighbour printed after it: ink onto bare shirt, and a later ink over an earlier one, since screen printing goes wet on wet in print order.
- Compensation. With a gain set, the dots are placed by model-based error diffusion, an idea from printer-model halftoning (Pappas and Neuhoff). Each candidate ink is judged by how the cell will really look next to the dots already placed, and the ink it spills onto them counts as error already spent. With the gain at zero you get the same dots as before, cell for cell.
- A way to measure it. The page downloads a calibration strip per ink: a solid square, a 50% checkerboard of single dots, and bare shirt. The printer runs it on a proof. You pick three colours from one photo of it, and the page solves the model for the gain. In a checkerboard every bare cell touches four dots, so the inked share is 0.5 plus gain/200. It warns when the checkerboard comes out lighter than half, which means the mesh is dropping dots, not spreading them.
What it fixes
A grey ramp printed in black dots on a white shirt, blurred as an eye would see it, is 0.89 delta E from the original with no gain. Spread the dots by 20% and it drifts to 3.39. Compensated, it is 0.78. At 50% gain the uncompensated print is 9.12 off and the compensated one 0.62. For one-ink prints and anything tonal, the setting does its job.
What it doesn't
Then we ran the same nine public-domain paintings as this morning (Van Gogh, Munch, Turner, Hokusai, Matisse, Picabia, Sassoferrato, Barocci) with four inks, and rendered each print at a true gain of 15% or 30%. The averages at 1 m:
| Shirt, true gain | No gain | Plain dots | Compensated | Compensated for double |
|---|---|---|---|---|
| white, 30% | 10.64 | 10.53 | 10.40 | 10.76 |
| near-black, 30% | 11.04 | 10.91 | 10.86 | 10.88 |
Nothing moves much. Four inks cannot mix many of a painting's colours, and that gap of 10 or 11 delta E swamps everything else. Under our model the spread even helps plain dots a little, because it blends them. Compensation is worth 0.1 to 0.26.
Two findings carry over to a real press:
- Guessing high is worse than not compensating. Compensating for double the true gain beat plain dots on only 2 or 3 of the 9 paintings. Compensating for half the true gain did as well as the truth (9 of 9 better at 30% on white). If you can't measure, guess low.
- For paintings, spend the effort on the inks. Choosing inks for the dither, not for flat colour regions, is where the next gain is. Dot gain compensation matters more once those colours are within reach.
None of this has been on a real press. The spread model is a first-order guess: no ink trapping, flash curing or squeegee direction. The calibration strip exists so a printer's proof can show whether it holds. The method and every number are in the studio's README under "Dot gain", and the measurement script is eval/index-gain.mjs.