We gave it more material, and the comic got worse

When a chapter ends, you can now turn it into a comic: a few pages of panels with speech balloons and captions, downloadable as a single PDF.

Here is how it works. The model gets the chapter’s summary and the cast list, and returns a storyboard — pages, panels, and for each panel the shot, the camera, who is in it and what is happening. Each panel is then painted separately. The dialogue and sound effects never go into the image at all; Pillow letters them on top afterwards. That is not a style choice, it is a constraint: the image model cannot write, and everything it tries to write comes out as a pattern that merely resembles text.

Painting costs real money, so the storyboard step runs first and on its own. It is cheap, and everything stops there: you get an estimate, and not one panel is painted until you confirm it.

The obvious idea

The summary is short — one paragraph. A comic made from it comes out at a single page, and for a chapter that visited three or four different places, that is not much.

Meanwhile the database holds the full play log: every action the player typed, every reply the Master gave. Many times the material. The reasoning writes itself — use the log instead of the summary and you get something longer and more detailed.

On 23 July we did exactly that. On a real, played chapter: the summary is 1,154 characters and produced one page. The log is around 5,700 characters and produced two pages and eight panels.

Twice the pages. A worse comic.

What actually broke

People started appearing in establishing shots who were not in the scene. The log mentioned the smith in a neighbouring line, and the model put him in the frame. Characters ended up somewhere other than where they had been placed two lines earlier. And the panel descriptions got noticeably denser, which is where the image model started tripping: the more detail you hand it, the more happily it duplicates props and paints in things nobody asked for.

We went back to the summary.

Why it went that way

We had been treating the summary as loss — a long text squeezed into a paragraph, with some of the detail thrown away. It turns out the throwing away is the work.

The summary hands the storyboard a handful of clear scenes. The log hands it a turn-by-turn record with no scene boundaries marked, so the model has to cut it into scenes itself, deciding as it goes what was a separate moment and what was a continuation of the last one. It does that, and it does it badly — failing in exactly the places you would fail if somebody gave you a transcript where you had asked for a script.

More input is not the same thing as more useful input. We knew that in the abstract and still spent a day proving it on our own material.

What comes next

The right way to get the log’s detail without its noise is not to swap the source but to work in two passes: pull the scene skeleton out of the summary first (which scenes, who is in them, where they happen), then read the matching stretch of log for each scene and sharpen the detail there. The clean scene decomposition survives, and the detail lands inside panels that have already been laid out.

That is not built. It is written down as a direction, not as this week’s plan.

And in fairness, of what is known and still open: panels draw characters in the wrong gear. Appearance goes in as free text, and the equipment actually worn never reaches the prompt at all. So your hero can walk into frame without the sword you just bought him.