Twenty-Three Out of Twenty-Eight Drew a Maple Leaf

AICU Japan ran a contest to mark the release of its AICU Create tool: one vertical image on the theme of making art in autumn, submitted between October 8 and 20. Thirty-one entries arrived from nine countries.

The report the organisers published afterwards is more useful than most contest write-ups, because it includes a tally nobody usually prints. Of the 28 works they counted for tendencies, 23 contained maple leaves.

That is not a knock on the entrants. It is a measurement of how a brief behaves when a generative model is asked to satisfy it. Autumn plus art plus Japan resolves, for a diffusion model and for most humans, to a girl with a brush, red foliage, and warm light. Ask a thousand people and you get the same picture with different faces.

The spec that mattered more than the theme

Buried in the entry requirements was a constraint that separates a picture from a cover. Entries had to be B5 portrait, with title space reserved at the top and bottom of the frame. The final file, if selected, needed to be 2000 by 3000 pixels.

That is a layout instruction disguised as a format. A magazine cover is not an illustration with a logo dropped on it. The masthead has to sit somewhere, the cover lines need a dark or quiet region to live in, and the subject cannot put its most important detail exactly where the type will go.

The organisers said so directly when describing the shortlist: they were looking at both the appeal of the image and its usability as a cover. One submission got credit for a moonlit atelier where the star field at the top and the dark tatami at the bottom left clean space for text. Another was praised for stacking depth in a way that kept the figure legible at a glance, with large sheets of paper in the foreground and smaller ones behind, so the eye has somewhere to travel.

Read the shortlisted descriptions as a list of what actually scored, and a pattern shows up. Overlapping shapes that frame a face. A single strong line, like a gold crack running across an ultramarine sky, that leads straight to the brush tip. Repetition of one accent colour rather than five competing ones. A foreground element, a sleeping dog or a pot of brushes, that gives the frame a near plane.

What converged, and what did not

Almost everyone reached for the same two or three solutions: a girl holding a brush, a palette, a window with autumn outside. The organisers noted that drawing with brush and canvas was by far the most common interpretation of "making art."

The entries that stood out changed the instrument or the vantage. Some showed characters playing guitar, bass or harp. Someone was assembling a pumpkin robot. Someone was spraying a wall. Someone was painting a lantern leaf by leaf. One illustrator from Brazil pressed ginkgo leaves into paint and used the leaf itself as the shape, which is a physical technique translated into an image rather than a description of a technique.

None of those are radical ideas. They are just the second idea instead of the first. That is the whole trick, and it is hard to do on purpose because the first idea arrives instantly and looks fine.

Where the time actually goes

The deadline data is its own small lesson. Eighteen of the 31 entries arrived in the final three days, and 11 arrived on the last day. The contest ran thirteen days. Most people did the work at the end, which means most people were generating against a closing window with no time to reject their first concept.

If you are producing images for a real cover, the order of operations matters more than the prompt. Decide the type placement first. Decide what the image has to say at thumbnail size, because that is how a cover is actually encountered on a shelf or in a feed. Then write the prompt with those two facts inside it: composition that leaves the upper third quiet, a subject that survives being 200 pixels tall, one accent colour.

Most AI illustration advice inverts this. It starts with style words and lets composition be whatever the model decides. Style words are the least scarce thing in the pipeline now. A model will give you a beautiful rendering of a badly composed frame all day, and it will not tell you the masthead has nowhere to go.

Three constraints worth putting in the prompt

Reserve the type zone in the composition, not in post. "Vertical composition, dark quiet sky occupying the top quarter, subject in the lower two thirds" does more for a cover than any quality tag. If you crop the type space in later, you are cropping the picture you designed.

Pick one motif and make it do work. Those 23 maple-leaf images are not a failure of imagination so much as a failure of editing. A single repeated element, used as a structural line or a colour, reads as intent. Five seasonal symbols read as a stock photo.

Check the image at 15 percent. Open it small enough that detail disappears. If the subject is not readable and the composition is not clear, the illustration is not a cover, whatever the resolution says.

Where this stops being general advice

A contest brief is not a client brief. This one was built around a specific magazine, a specific trim size, and a specific prize, and the judges were editors who needed a cover they could put type on. That bias is the reason the observations hold together, and also the reason not to over-generalise them. A poster, a book jacket and a thumbnail for a video all reserve space differently, and a cover for a print run has constraints a web hero does not.

What transfers is the habit: find the destination format before you generate, put its constraints into the prompt in plain language, and treat the first idea as a draft that needs to be replaced. The entrants who did that are the ones whose images the report described in detail instead of listing.

If you are working on character-driven covers rather than single illustrations, the text-first character method is the better starting point, since it fixes identity before composition. And when a cover has to work across several placements, the multi-format pipeline notes cover the crop decisions that come after the image exists.