🎫 concert-series

I built a flyer generator. It reads a folder of YAML, one file per show, and writes an 8.5 by 11 inch SVG. The page bisects, a photograph fills one half and bleeds off the trim, the type sets in the other half on a grid, one ink on a coloured ground. The photographs are screened as newsprint halftones.

Then I asked an image model for the same six flyers. Same content, same brief, typed into a prompt instead of parsed from YAML. The ground colour and the ink were the ones the generator had already assigned, so method was the only variable.

The diffusion flyers are good. All six got the register: flat ground, one ink, halftone photograph, heavy grotesque, a defensible hierarchy. Two of them I would accept from a designer.

Then I measured them.

The page is 8.5 by 11 inches. Every image came back 1024 by 1536, which is 2:3, the nearest portrait size on offer. There is no parameter for trim size. The brief asked for one flat ground and exactly one ink; the six images hold between 61,000 and 185,000 distinct colours, and in the worst one the commonest single colour covers 0.6 percent of the page. ENNIO MORRICONE runs under the photograph it was told to sit clear of. Two prices lost their dollar sign and a third went missing. One flyer typeset the words “Opening acts” because my prompt used that phrase as a field label.

None of those is a failure of taste; they are all the same failure. The model is painting a picture of a flyer. It has no page, no measure, no colour value, no string that has to survive intact. It has a very good idea of what those things look like.

The generator has all of them. It knows the page is 8.5 by 11 because the page is a number. It knows DE ANGELIS fits because it read the font’s metrics table and added up the advance widths, and when a name does not fit it shrinks that line until it does. Run it again and you get the same bytes.

That argument is old, and procedural design has always lost it anyway, because writing the system cost more than drawing the thing. The generator parses TrueType metrics tables from the spec, because I did not want the dependency. It resolves every unspecified slot deterministically, sets a column grid, and screens each photograph at the size it will print. I did not write most of it. I specified it, in conversation, and corrected it when it went wrong. The expensive half of procedural design got cheap.

The clearest demonstration was a mistake. Halfway through, the halftone looked mechanical. The screen was sitting at 45 degrees, which is the classic angle for a single ink and also the one angle where a square screen lands every dot back in a straight row. I changed one number to 22.5 and the whole series was right. With the prompt I would have rolled again and taken what came.

Diffusion is the right tool when you want one image, you will not need it again, and nothing about it has to be true. Plenty of work is like that.

I could not have written that prompt before I built the generator, though. The halves, the bleed, the single ink, the gutter against the photograph, the screen ruling, flush left on a grid: I knew none of it when I started. The build argued me into it, one wrong version at a time. The prompt is the design document the generator produced, read back out, and the diffusion flyers came out well on borrowed work. Someone writing that brief cold would not get close.

The generator was the quick part. What took the time was working out what it should do, and that is the part worth keeping. Building the system forces those decisions, because every vagueness comes back as a crash or an ugly page. The thinking ports to any toolchain, including the one that paints pictures.