AI Story Illustrations Are Better One per Story

AI story illustrations are most useful when there are almost none of them. By the time a picture of the room appears, you have already built that room out of six or seven details in the prose, and the image’s job is no longer to show you something — it is to overrule you. One illustration per story, drawn late and from a scene you actually reached, leaves the reading intact and still gives the story a face.

• Inform, the long-established system for writing interactive fiction, calls pictures “figures” and describes a work of IF as a mostly textual book broken up with illustrations here and there.
• In a released Inform story, resource ID 1 is reserved for the cover art and figures are numbered upward from 2, so the cover is formally a different kind of picture from the illustrations.
• A figure shown mid-story scrolls away like text; Inform’s documentation calls such pictures part of the stream of narrative.
• Images that are not preloaded arrive late, which is why Vorple’s multimedia extension tells authors to preload them before the scene that needs them.
• Alt text for a story image has to give the same information to a reader who cannot see it, which is a useful test of whether the image was carrying information at all.

Why do AI story illustrations compete with your own picture?

Because they arrive second. Prose hands you a small number of fixed details — the color of the light, one sound, the fact that a door is half open — and you supply everything else without noticing: the height of the ceiling, the wear on the floor, the face. An image arrives into a room that is already furnished, and you do not get to keep both versions. The one on the screen wins.

That is a claim about order rather than about pictures. A cover painting arrives before you have built anything, so it overrules nothing; it sets a key, the way a dust jacket does. A picture dropped into scene forty arrives after four hours of work you did yourself, and most of that work it quietly contradicts.

The cost is highest for faces. A character you have read for forty scenes has a face assembled from a gesture, a way of not answering questions, and one description you half remember. Any generated portrait of that person is a stranger. The mismatch is not a matter of taste — it discards something you made, and it cannot give it back.

A form written in second person and present tense leans on that private construction harder than most. “You push the door” only works if the door is yours. The more of the room you are supplying, the more a picture has to lose.

What text adventures already settled about figures and covers

The interactive fiction tradition worked this out decades ago and encoded it in the tools. Inform’s own documentation, “Writing with Inform”, says Inform “calls these pictures figures, following the usual Inform analogy with books”, and asks the author to think of a work of IF “as being like a mostly textual book which in broken up with illustrations here and there — Figure 1, Figure 2, and so on”.

Two details in that system are worth borrowing. The first is that figures are transient: “Once seen, they scroll away, just as text does once it has been printed. These pictures are really part of the stream of narrative.” An illustration is an event in the reading, not a permanent panel.

The second is that the cover is not a figure at all. In a released story, the documentation notes, “Resource ID number 1 is reserved for the image of the cover art, if there is any”, with figures numbered upward from 2. Cover art is recommended square, “like a music album cover”, at 960 by 960, and is the image that interpreters and the IFDB show when they list the work.

So the format itself distinguishes one picture that stands for the whole thing from a handful that interrupt it. That is close to the opposite of the tradition where the image leads and the text follows, which is most of the difference between interactive fiction and a visual novel.

What one AI story illustration does that a gallery cannot

It can be specific. A per-scene generator has to draw the forest, because it is working from a paragraph and a genre. A single image drawn at the midpoint can be about your story — the mirror that showed what was behind it, the corridor you refused to leave — because by then there is a story to be about, with particular objects in it and particular decisions behind them.

Timing does most of the work. Draw too early and you have produced stock art for a premise; draw at the end and the reader has finished and is unlikely to look. The midpoint is the first moment at which an image can be both earned and useful, and it is also the moment at which a reader is most likely to want a keepsake of what they are in the middle of.

This is how we build it. Fabledrift, an interactive fiction app for Android, ships exactly two kinds of picture and no others: a shared cover painting for each world, and one illustration per story, drawn in the background at the story’s midpoint from the scene the reader has reached. Scenes are otherwise text. There is no per-scene image and no paid image feature. The two launch worlds and their covers — The Emberwood and Ironreach Station — carry the painting that a reader sees before the first scene; everything after that is prose and one picture.

What per-scene art costs the reading

Three things, and only one of them is money. The first is the mental picture, discussed above. The second is time: an image has to be generated or fetched before it can be shown, and the reader sits in front of an unfinished page while that happens. The third is consistency, because a character redrawn eight times is eight people.

The wait is not hypothetical, and the interactive fiction tools treat it as a known problem. Vorple’s multimedia extension for Inform documents preloading precisely because otherwise “the images are loaded only when they are first displayed which may take some time with slower connections, resulting in a noticeable delay between when they should be shown and when they have loaded and actually appear”. Preloading a file you already have is possible. Preloading one that does not exist yet is not.

Property	An image every scene	One illustration at the midpoint	Cover painting only
What it can show	The scene in general	This story in particular	The world’s key
When it arrives	Mid-scene, every scene	Once, in the background	Before scene one
Wait before you read on	Each time	None	None
Consistency risk	High, and cumulative	Contained to one picture	None
Effect on your own picture	Replaced repeatedly	Replaced once, late	Set before you start
What the reader keeps	A gallery of near misses	A portrait of their story	A shared image

The alt text problem

Accessibility is also a test of substance. Vorple’s documentation notes that an image description “is shown in standard interpreters instead of the picture and read aloud by screen readers, so it should be a compact but accurate description of what the image depicts”, and that where seeing the image matters, the description “should give the same information to those who can’t see the image”. Inform allows the same short text on any figure and on the cover.

Try writing that sentence for a per-scene illustration. It will come out as a paraphrase of the paragraph the reader just read, because that is what the image was generated from. When the honest alt text is a summary of the adjacent prose, the picture is decoration, and it is competing with the reading for nothing.

The picture that arrives after the room is built

There is a simple test for any story illustration, and you can run it after you finish something. Ask whether you remember the image or your own version of the place. If the picture agreed with what you had already built, it earned its place. If it replaced it, the story is now shorter by one room you made yourself.

That is the whole case for restraint with AI story illustrations. Not that pictures are a lesser medium, but that in a story made of sentences they are expensive in a currency nobody bills for: the reader’s share of the work.