The comic generator is the oldest of the creative projects in this estate and the one that has changed shape most often. The earlier post in this section is about one bug in one schema. This piece is the long view: what the generator was for, what each version measured, where it stands in October 2026, and which architectures were kept and which were dropped. The people in the books are named by role: the mother, the father, the kids, the grandmother, the grandfather, the uncle, the aunt, and two villagers from the game’s world who get a cameo.
The goal
The first goal, in June 2026, was narrow: a story prompt and a few photographs go in, and a kid-appropriate comic comes out, with the characters drawn from real people and kept consistent across every panel, variable panel sizes, several pages, speech bubbles that read as speech bubbles, a front and back cover, and a PDF and CBZ at the end. All of it had to run on hardware already in the house, first a laptop with unified memory, later a desktop with a 16 GB card.
By October the goal had grown a second half. The comic stopped being a stand-alone generator and became the game’s companion: a sixteen-page book set in the game’s village, drawn from its own rendered scenery, with the family as the cast, printed and stapled at home. “Done” is written down as a PDF and a CBZ that the playtester approves with no hand step left after one colour choice. And behind that sits a third goal that is still being built: a studio where someone who is not the author can make a book without a terminal.
What we tested, and what the counts said
The slicer. The first pipeline put everything inside a node-graph image tool. A language model in a desktop app wrote a multi-panel script in one pass; nodes in the graph split the text on delimiters into art prompts and dialogue, painted each panel in turn, stamped vector text over the art, batched the panels and stitched them onto a grid with 15-pixel gutters. It made pages. It also produced a failure log that is still worth reading: a ||| delimiter that silently crashed the regex, a ### one that Markdown ate, a model that forgot who the characters were unless every panel re-stated them, a stretched canvas fixed by locking panels to 832×1216, a divide-by-zero when a face detector found no face, and a threading crash in a plotting library. A reasoning model exhausted its token budget formatting the script, so a 12-billion-parameter instruct model was adopted for speed and a 32-billion one for consistency. Within a week the diagnosis was that nearly every entry in that log was control flow, layout or typography forced into a dataflow graph.
The Director. The rebuild moved every loop, every validation, the layout, the lettering and the packaging into a thin Python program and left the graph one job: paint one textless, reference-conditioned panel. The writer now emits schema-locked JSON, a comic spec that is the contract between stages. The same week, three polish tracks landed with measurements: a writer prompt that pushed characters mid-action because the earlier scripts read as talking heads, face-aware captions that flip bands when their own band would cover a face, tail-less bubbles when one speaker’s head fills 12 % of the panel, and a bubble font that re-plans smaller when a five-bubble panel overlapped by about 33,000 square pixels at the fixed size. The suite stood at 300 tests then. A full book in a new genre, 21 panels and two covers on the same two adults from the household, exposed the limits that still hold: the quality gate scored a three-armed figure 1.00 and marked an undefined bandit as a duplicate head, and an unnamed side character made the painter clone the hero.
The silence. When the schema gained mood and settings fields, the writer stopped writing dialogue and nothing in validation noticed, because every field was optional and present. That story has its own piece; the measurement after the fix was three talking books out of three on a mixture-of-experts writer, and the pattern of relocating fields out of the schema and deriving them from prose became the rule for every cross-panel registry.
The long book. On 1 July 2026 the pipeline passed its long-book gate: fourteen pages, 56 panels, 87 lines of dialogue, identity held across every panel, ten panels flagged by the quality gate, in one hour and 47 minutes at a median of 115 s per panel. A stability review by five independent reviewers followed, and the remaining open issues were counted at four. That book became the fixture for everything after it.
The composed book. In October the comic turned to face the game. Instead of painting scenes from prompts, a camera script in the game engine rendered plates of the village, the bay, the meadow and the inn, and the painter repainted each plate in edit mode with the plate as its only reference, which kept the composition and added realism. The first booklet was eight pages, hand-duplexed on a home printer; seventeen plates at about 30 s each. The second was twelve. The third, planned on a storyboard page before anything was rendered, found the defect that had been hiding in both: the figures were 2.2 to 3 times too large for their plates. The plates had been painted empty, the character art was height-normalised, and the composer typed a fraction of the panel height with no link to the camera that made the plate. The fix was to project each figure through the plate’s own camera and add a proportion gate, 10 % on a figure and 5 % on a pair, with real heights on file for the whole cast from 88 cm to 183 cm. Prompt wording ranked last among the candidate fixes.
The print path. The third book also brought the print stage: oval balloons, a 300 DPI path to a US trim size with bleed, imposition for a saddle-stitched booklet on three sheets, and a test strip because the cover was 44 % out of gamut on the press profile and the end page 52 %. The build painted 140 character cells at about 11 s each, composed sixteen pages with eleven stock fallbacks, and stopped at two test strips for the playtester to choose from.
Where we are
The fourth book exists as sixteen pages, a reader PDF, a print PDF and a CBZ. Its review pages were redone after the first review showed only four of them: a review that shows a subset reads as an incomplete book, and every review since shows every page with a verdict. The family saw the first web preview in mid-October and enjoyed it, which fixed the shape: the parents and the kids front and centre, then the grandparents, the uncle and the aunt, and the two villagers at the edges. The colour choice and the final export still wait on the strip.
The pipeline has three render backends behind one switch, including one that runs a four-billion-parameter painter in 2.6 GB of video memory with sequential offload. The writer is a 35-billion-parameter mixture-of-experts model behind an outline-then-expand pass. The studio has moved from a nine-screen rehearsal over a mocked event stream, 3.8 thousand lines of React against the real event contract, to its first four slices merged: a login door, a blob store behind one interface, a job queue that uses the farm’s capability pickup, and the screens that were approved from mockups on 19 October. The suite reports 985 tests passing and 14 skipped.
On 21 October the guided-panel reader the family previewed became a project of its own, shared by the generator, the studio and the review pages, with a headless test that the panels move in the right order on a fixture book that needs no family art. The studio was brought up on the desktop for the first time, which found a missing imaging library and three configuration paths, and every book route now checks the project’s visibility before it serves a page.
What is next
Pick the strip, run the colour transform, export, and close the fourth book. Make the print stage write the CBZ itself instead of a hand step. Build the studio’s remaining slices against the approved screens, with owner, editor and viewer roles per project and likeness projects private by default. Add a wardrobe registry read from the prose, because a coat still vanishes between panels. The forge now owns the painter and the cast, so the comic’s next pictures depend on the forge journey as much as on this one.
Architectures we tried
Kept: a code Director over a stateless painter; schema-locked JSON with fields derived from prose rather than demanded; content-addressed panel variants with provenance; a composed book from game plates with a proportion gate; a web studio over a server-sent event stream. Dropped: orchestration inside the node graph, with its delimiters, safety valves and batch tensors; a reasoning model as the writer; prompt wording as a fix for proportion; a four-page review. Considered and not chosen: a Python-native UI framework for the studio, and putting the studio inside the admin console, where one door would have excluded everyone who is not an administrator.