← The Lab

Methodology

How we build comic styles: a contract, not an adjective

Most AI "styles" are a mood word in a prompt. Ours are studied from public-domain plates and distilled into rules a machine must follow — and a test can fail.

Chess2Story Research ·

Ask any AI comic generator for "noir style" and you will get rain, a trench coat, and venetian-blind shadows — painted in soft digital greys, with a glowing streetlamp and chrome-bevelled lettering. The subject changed; the drawing didn't. That is what a style is in most tools: an adjective that redirects the model's attention while every actual mark on the page stays the model's house default. Style references and LoRAs push harder than a suffix does, but they still only nudge — nothing measures the page that comes out.

A Chess2Story style is not an adjective. It is a technique contract: a written specification of how a mark is made in one historical tradition, derived from public-domain plates, enforced in the production prompt, and written as measurable tests a style's output is reviewed against.

Why AI-generated comic art looks generic

Because a model fills every unspecified detail with its statistical default: smooth gradients, airbrushed shadows, glowing light, bevelled lettering. A prompt that names only a mood leaves everything unspecified, so the default takes over — which is why "noir," "woodcut" and "watercolour" all come back looking like the same soft digital render wearing different subject matter. The fix is not a better adjective. It is specifying the construction — what a mark is, how light is made, what the lettering idiom is, what is forbidden — and then measuring whether the finished page complied.

Here is the pipeline, using our Noir style as the worked example — every number below is a real measurement from Noir's research dossier, quoted with its source named.

Félix Vallotton, Les Nécrophores, 1892 — relief print, public domainNoir's reference plate: Félix Vallotton, Les Nécrophores (1892). Measured 44.5% black, 52.3% white, 0.0% flat grey on the file we ship.

Step 1 — harvest evidence, and let it say no

A style starts as a research question: what physical artifact is this tradition, and what would prove a page belongs to it? For Noir we collected twenty-six candidate artworks across three rounds — Vallotton relief plates from 1892–97, four of Beardsley's Salome drawings (1893–94), Expressionist woodcuts by Munch, Kirchner and Kollwitz, plus paintings, etchings, other prints and film stills tested as controls.

Sixteen survived as first-class evidence, and two more were demoted to subordinate witnesses. Eight were rejected outright, with written reasons — and the rejections teach more than the keeps. Of the five Expressionist film stills we tested as the tradition's cinematic ancestors — Nosferatu and Caligari frames — three were rejected, and the two that survived are exactly the two subordinate witnesses: one kept for the geometry of its shadows, one for the shape of its letters. Neither was allowed to teach a mark. Everything a cinematographer manufactures (falloff, fill light, bloom, grain) is a tonal ramp, and this printmaking tradition has no ramps. Film could donate the rain and the wardrobe, but not one mark.

Every plate in every evidence base is public domain or CC0 — Noir's reference plate is on Wikimedia Commons. Living or in-copyright artists are named as lineage — naming a tradition is free — but their images are never fetched or used.

Step 2 — extract rules a machine can obey

From the surviving plates we write the style's definition: not "moody and dramatic" but testable construction rules. Each rule has to be countable. "Faces work in exactly two systems" is a tally, not a vibe: a bare-paper face carrying four to six black marks and no outline, or a face carved inside the black as one white gouge with the eyes and mouth left standing as small black islands. Vallotton's L'Exécution (1894) shows both systems on one sheet — which is why the definition names two systems, not one rule with an exception.

The same discipline produces the rest: black is poured as unbroken hard-edged shapes and every white is a removal; the page is bimodal, not dark — black coverage across the canonical plates runs 17–79%, and what never varies is that everything resolves to ink or paper; light is a cut shape with a hard border, never a drawn source.

The distinctions between neighbouring styles are designed here too, one axis per neighbour — a Noir black is empty while a Gothic black is built from countable strokes — so each style in the fleet stays itself instead of blurring into its neighbours. It's why a king hunt in Noir looks like a black mass eating the board, and the same king hunt in Ukiyo-e looks like a crowd of figures stacked at one scale, nobody shaded, nobody glowing.

Step 3 — write the prompt as a technique contract

The production prompt never says "dark," "moody," or "high contrast" — phrases a model hears as greyscale, which is a continuous medium and therefore the exact failure. It specifies construction, in order: flood the black first, then cut the light out of it. Positive build rules make the failure structurally impossible — you cannot feather a gouge.

Three more artifacts travel with the prompt:

  • A composition contract — how splash, dialogue, and action beats are staged in this tradition's grammar.
  • A lettering idiom — Noir's caption box, a white rectangle with a black keyline, appears four times in Vallotton's own plates. Period ancestry, not a comics import.
  • A reference plate the model can see — for Noir, Félix Vallotton's Les Nécrophores (1892, public domain), which measures 44.5% black, 52.3% white and 0.0% flat grey on the file we ship.

Step 4 — gate the output with numbers

A finished page is judged against acceptance tests calibrated on the historical plates. Noir's hardest gate: flat middle-grey may not exceed 5% of the page. The sixteen canonical plates measure between 0.00% and 0.13%. Six of the eight rejects land between 7.1% and 43.6% — an order of magnitude clear of every keep, with nothing in the gap. The other two slip through at effectively zero, and they are the reason this is not the only test: their tone is grainy rather than flat, which a flatness measurement is blind to by construction, so a second diagnostic and a human check at zoom catch what the gate cannot. A single coloured pixel — including the cream caption boxes most digital comics sneak onto "black and white" pages — is a hard fail on its own.

What this doesn't solve

Honesty is part of the method, so here is what step 4 is and is not. Generation is probabilistic: a model sometimes ignores its references, and a gate describes that drift rather than preventing it. The style measurements above are thresholds derived from the plates and the bar a page is judged against — they are not wired into the renderer as a per-page pass. What does run automatically on every generated page is narrower: a vision check that transcribes the lettering on the finished art and diffs it against the page script, looking for a doubled caption, an invented one, or a reference card pasted into the picture — and even that one is advisory, since a checker that is unavailable never blocks a render. When the retries do run out, the page ships with its remaining lettering problems recorded rather than hidden. Claiming more than that would be the exact move this essay was written to argue against.

The write-ups also lag the research: all twenty-one styles already have a research dossier behind them — a harvest log, per-plate keep-or-reject verdicts, a monograph, gate thresholds. Fourteen have a public write-up so far, Noir and Ukiyo-e among them, with named plates, collections, accession numbers and thresholds you can check. Until a style's write-up lands you can see its rules in the output but not the plates behind them.

The proof was never going to be this essay. It's the pages: browse the published comics, pick any panel, and run the tests this article describes — no feathered shadows, no glowing light sources, no lettering the tradition wouldn't cut, whatever style the comic declares. That's the contract, and you're the inspector — and when you're ready, your own game can be next.

Frequently asked questions

How are AI comic art styles usually made?

In most AI comic tools a style is a short prompt suffix — "noir style, dramatic shadows, high contrast" — sometimes with an artist name attached. Better tools condition on a style reference: a LoRA, an IP-Adapter image, a Midjourney sref code. What almost none of them add is the third piece: a written definition of what the tradition's marks actually are, and measurable tests written from the plates, which a style's output is reviewed against. Reference-conditioning nudges the output; it does not measure it.

What makes Chess2Story's comic styles different from other AI styles?

Four layers. First, an evidence base: public-domain plates collected in documented research rounds, each with a written keep-or-reject verdict. Second, a definition: the tradition's rules extracted from those plates as testable statements. Third, a technique-contract prompt: instructions for how marks are constructed, not how the page should feel. Fourth, calibrated gates: measurements tuned against the original plates, which are how a style is validated and its output reviewed for drift toward generic digital rendering.

A style — a way of making marks — is not copyrightable; specific images are. The legal question is what images the system actually ingests. Chess2Story's evidence bases contain only public-domain and CC0 images: 1890s relief prints, Edo-period woodblock sheets, Expressionist woodcuts. Artists still in copyright are cited by name as part of a tradition's lineage — naming a tradition is free — but none of their images are ever fetched, stored, or used as references. The provenance of each researched style is published in its Lab article.

Judge the output, not the pitch

Before you spend anything, look at finished pages. Then paste a game of your own.