A good concept poster is not a checklist arranged on a page. It is a sequence of discoveries: the title catches you, the image changes its meaning, and one small detail tells you how to feel about the whole thing. Image generators can produce that sequence, but only if the prompt describes relationships instead of piling up style words.
We tested that idea with The Night Garden, a fictional exhibition built around a giant botanical moth and a tiny lamp-carrying botanist. The experiment produced four useful images: one attractive failure, two different posters from the same detailed prompt, and one repair that fixed the composition while quietly changing the moth. Together they show where a prompt helps, where it stops helping, and what a human still has to notice.
The attractive version failed first
Our first request asked for a striking vertical retro poster, a giant moth, a very small botanist, elegant print texture, dramatic scale, and the title The Night Garden. It sounded specific. The result looked expensive enough to survive a quick review: a pale serif title floats over a dark garden, a moon lights the upper corner, and the oversized moth immediately creates atmosphere.
Then we read it. The image had written three extra pieces of exhibition copy that nobody requested. It added flowers, water, moonlight, and a complete nocturnal landscape. The “print texture” became a photographic vintage treatment rather than a flat editorial print. The poster was visually coherent, but it belonged to a different brief.
That failure changed the way we wrote the next prompt. More adjectives would not have helped. What was missing was a decision about what the viewer should notice first, where the page had to remain empty, how the title should meet the moth, and which words were allowed to exist.
Five relationships make the poster
1. Reading order. A poster is read in time. For Night Garden, we wanted the title first, the moth second, the tiny botanist third, and two smaller lines last. Writing down that sequence gives every later choice a job. A bright lamp is useful only if it does not steal the first glance; supporting copy is useful only if it waits its turn.
2. Image and type. The title should not look as if it was pasted above an illustration. We gave the moth permission to cross the lower edges of a few letters, but asked that at least 90 percent of every letter remain visible. Think of this as an occlusion budget: enough overlap to make the two layers interact, not enough to turn reading into a puzzle.
3. Dominant form and scale anchor. The botanist was set at roughly one twelfth of the moth's height, in profile at the lower right, holding one lamp and looking up. Those details keep the person from becoming a second protagonist. The figure's familiar silhouette makes the moth feel enormous before the supporting text explains anything.
4. Subject and quiet space. We reserved the top 22 percent of the warm-ivory page as empty paper. No wing, word, border, or knot of grain could enter it. This was not leftover space. It was the pause that made the title and moth feel monumental instead of crowded.
5. Idea and material. “Retro” can mean almost anything, so we replaced it with visible behavior: flat shapes, direct line drawing, uncoated ivory paper, fine halftone, restrained ink wear, and tiny registration shifts. We also named what did not belong—photorealism, glossy gradients, cinematic depth, polished 3D, and artificially yellowed paper.
| Vague direction | Something you can actually inspect |
|---|---|
| Make the hierarchy strong | Title first, moth second, figure third, small copy last |
| Big moth, tiny person | One person at about one twelfth of the moth's height |
| Blend the image and title | Small overlap at the lower letter edges; every word remains immediately readable |
| Leave plenty of space | The upper 22 percent stays completely empty |
| Use a retro print style | Named shapes, line behavior, paper, ink, halftone, and registration |
The same prompt did not produce the same poster
Run A and Run B came from the exact same filled prompt. Both kept the three approved lines of text, the empty upper field, one moth, one botanist, one lamp, and the navy-green-orange palette. Both also preserved the intended first impression: the page feels quiet, and the person is visibly dwarfed by the insect.
They are still different designs. Run A builds the moth from several leaf-like lobes and keeps both smaller lines in navy. Run B uses a cleaner four-wing silhouette, larger orange spots, and a green final line. The spacing around the botanist changes, as does the distance between the title and the moth. The prompt held the design logic; it did not freeze the drawing.
The shared weakness became obvious when the images were viewed together. Run A lets the antennae touch the title area, while most of the wings remain safely below it. Run B brings the upper wings closer, but the headline still reads more like a separate layer than part of the illustration. Both posters were usable examples, yet neither fully delivered the relationship that made the concept interesting.
The repair solved the headline and changed the moth
We used Run A as the edit source and asked for one visible change: raise the moth until small portions of both upper wings passed in front of the title. The prompt repeated the exact three text strings, the empty top area, palette, margins, figure, lamp, and lower copy as things to preserve.
The edited poster has a stronger headline. The wings now press into the bottoms of NIGHT GARDEN, while the words remain easy to read. It feels less like an illustration under a title and more like one composed surface.
But the image did not simply move upward. The body became narrower, the upper wings lost some of Run A's layered leaf construction, and the lower pattern changed. That is the important production lesson: “change only this” is a request, not a guarantee. A generative edit can solve the relationship you named and redraw something you hoped to keep.
CHANGE ONLY
[Describe the one relationship that failed and the visible correction.]
PRESERVE
[Restate the approved text, counts, composition, subject identity,
palette, crop, quiet space, and every detail that still works.]
DO NOT
[Name the collateral changes and additions most likely to occur.]
Return one corrected flat poster only.Use the template as a brief, not a form
The copyable Prompt on this page asks for a primary idea, one dominant metaphor, a familiar scale anchor, a reading order, an exact text allowlist, protected quiet space, a bounded palette, visible material rules, and project-specific exclusions. Those fields are not there to make the prompt look comprehensive. Each one corresponds to a decision that changed our result.
If a field is empty, stop and make the decision before generating. A placeholder such as “[STYLE]” or “[EXCLUSIONS]” is not creative freedom; it is a question being handed to the model. The model will answer it, but probably without knowing the campaign, the brand, the viewing distance, or which mistake will cost you another afternoon.
For a personal art experiment, you can relax the controls and welcome surprise. For an event poster, cover, or campaign asset, complete the brief first. Download the poster prompt pack for the reusable template, the exact Night Garden inputs, and the repair pattern. The test record contains the saved dimensions, file hashes, and review decisions.
A six-pass review that takes minutes
- Read every character. Compare the image with the allowlist, including punctuation and small copy. Look around the edges for invented labels and pseudo-writing.
- Notice your first glance. Does your eye follow the planned order, or has contrast made a supporting element louder than the title?
- Inspect the contact point. Where type and image meet, check both the visual tension and the readability. Overlap is not automatically good composition.
- Count and compare scale. Confirm the subject and figure counts, then judge whether the familiar object really makes the dominant form feel larger.
- Look at the material. Check shape, line, surface, depth, and color as separate choices. A convincing mood can hide the wrong rendering method.
- Try the real destination. View the actual social crop, mobile card, print size, or cover thumbnail. A headline that works full-screen may disappear when the poster becomes small.
Do these passes separately. When everything is judged at once, a beautiful moth can distract from a misspelled line, and correct text can make a weak hierarchy feel finished. OCR may help find suspicious lettering, but stylized type still needs a person looking at the original file.
Know when to take the type out of the generator
Generated lettering makes sense when the copy is short and its interaction with the picture is part of the experiment. It becomes a poor production choice when the wording is legal, frequently revised, translated into several languages, tied to a brand font, or expected to align across a campaign.
In those cases, use the model to discover the subject, rhythm, empty space, and rough title placement. Then rebuild the final type in Figma, Illustrator, Photoshop, or another layout tool. You keep the surprising composition while making spelling, kerning, accessibility, and late changes deterministic.
What this adds, and where it stops
The X article by TaXue2025 supplied the useful starting idea: a concept poster can connect a dominant figure, a tiny scale reference, typography, open space, and print texture. This page does not reproduce that article's images, subjects, or prompt. Night Garden is a new test designed to find out which parts of that idea could become a reusable working method.
The broader GPT Image 2.5 Prompt Guide explains general prompt structure, exact text, references, settings, and editing. This Prompt is narrower: it is for editorial posters in which hierarchy and image-type interaction are the main problem. OpenAI currently positions Flare as the speed-first option and Sunburst as the quality- and precision-first option, but our built-in image tool did not identify the model behind these four files. They should not be read as a model comparison or a success-rate claim.
Percentages in a prompt are visual targets rather than pixel locks. Short approved text can still fail on another run. Color names and hex values do not create a print proof. The workflow also cannot settle brand, copyright, publicity-rights, or accessibility questions. Those limits matter, but they do not erase what the examples show: a more deliberate brief produced a more deliberate poster, and a focused edit still required a complete second look.
The useful rule is simple
Before generating, decide what the viewer sees first, what proves the scale, where the page breathes, how far the image may enter the type, and exactly which words are allowed. After generating, look at those relationships in the same order. The best prompt is not the longest one. It is the one that makes a creative decision visible enough for a person to accept, repair, or reject.
