06 · Posters and supporting images
At a glance
| Output | A set of images in one style |
| Time | 20 to 40 minutes |
| Type | Image |
| The hard part | Consistency, not generation |
The situation
An article needs illustrations, an event needs a poster, a product needs a set of icons or covers.
Anyone can generate one image. The value here is making a set look like a set.
The method: approve one, then batch
The classic failure of batch generation is drift — flat illustration, then photoreal, then a 3D render.
The fix is two stages.
Stage 1 · Fix the style
I need [N] images for an article about [topic].
Style:
[Be specific. For example: low-saturation abstract geometry in deep
green and cream, with a light grain; no human faces; no text of any
kind; generous whitespace; a single light source from the upper left]
Generate only the first one for now. Its subject is [subject].
Once I confirm it, reuse the exact same style description for the
rest. Don't improvise per image.Stage 2 · Batch
Once approved:
Style confirmed. Generate the remaining [N-1] with the identical style
description:
2. [subject]
3. [subject]
4. [subject]
Reuse the first image's style description verbatim; only the subject
changes. Before you return them, check: do these look like one person
made them?What a style description should cover
A reusable one covers six things:
| Aspect | Example |
|---|---|
| Colour | Low saturation, deep green primary, cream ground |
| Texture | Light grain, like screen printing |
| Composition | Generous whitespace, subject centred and low |
| Light | Single source, upper left |
| Abstraction | Geometric abstraction, not photoreal |
| Prohibitions | No faces, no text, no gradients |
The last row is the most forgotten and the most useful. "No text" especially — generated lettering is usually garbled and looks terrible in real material.
Acceptance
- [ ] Seen together, they read as one set
- [ ] No text appears in any image
- [ ] Colour matches your brand or the piece's register
- [ ] Dimensions suit the destination (covers, inline art, and social cards differ)
- [ ] No obvious AI tells (extra fingers, unnatural symmetry, mushy detail)
- [ ] Each image actually depicts what it was supposed to
Variants
A poster, i.e. one with words on it
Don't let an image model set the type. Do it in two layers:
Layer one: generate a background image. Pure visual, no text at all.
Style: [description]
Composition: [leave space for type — e.g. "top third open"]
Layer two: set the type over that image in HTML/CSS.
Headline: [text]
Subhead: [text]
Typography from the current design system.Then the words are real text — editable, selectable, and sharp in print.
An icon set
Build a set of [N] icons for [purpose].
Shared spec:
- 2px stroke, rounded caps
- 24×24 grid with 2px padding
- Monochrome, using currentColor
- SVG output, not bitmap
The set:
1. [meaning] — suggested form
2. [meaning]
...
Each must remain legible at 24px.From a sketch
When you can't draw but you know the composition:
This is a hand-drawn sketch. Generate a finished image following its
composition.
In the sketch:
- [left block] = [what it is]
- [central shape] = [what it is]
Style: [description]
Follow the sketch's composition. Do not recompose.Questions that come up
Every image looks different.
You skipped the two-stage method. Approve one, then batch, and say "reuse the identical style description".
There's garbled text in the image.
Add "no text, letters, or numbers of any kind" to the style. Regenerate rather than trying to repair.
The images have nothing to do with the article.
Same test as step 6 of the deck workshop: an image explains a point or it's decoration. Decide what each image is saying before you generate it, and write that in.
It's going to print.
Say so up front: "final use is A3 print at 300 DPI — generate at the corresponding pixel dimensions and leave bleed."
Related
- Deck workshop / Step 6 — when a slide earns an image
- Go deeper / Media — choosing between images, motion, and video