Image generation basics
Worked case: from vague to art-directed
A strong final prompt is the trace of decisions made against a real brief, not merely a longer version of the first idea.
Before you start
Why this matters
The assignment
Imagine a public library needs a website hero for a free evening workshop called “Repair, Reuse, Share.” The image will sit behind a headline and registration button. The audience includes adults and teenagers. The library wants a hopeful, practical tone rather than luxury advertising.
The initial request is:
Make a nice image of people fixing things in a library.
This is understandable but not operational. “Nice,” “people,” “things,” and even “library” leave too much unresolved. The prompt says nothing about hierarchy, style, layout, representation, or exclusions.
1Learn the idea
Read
Baseline output diagnosis
Suppose the first batch produces polished images with these recurring problems:
- Six or seven people create a crowded scene.
- Shelves, posters, and glowing lamps compete with the activity.
- A large generated sign contains misspelled workshop text.
- Tools and hands are partially hidden.
- The image is square, but the site needs a wide crop.
- The most prominent object is a sewing machine, making the workshop seem sewing-only.
- One candidate resembles a recognizable electronics brand advertisement.
The mistake would be to append “better, cleaner, professional, high quality.” Those words do not address the failures. We need to reconstruct the brief.
Read
Step 1: clarify subject and story
The workshop needs to communicate collaborative repair across categories. Limit the scene to three participants and choose visible actions:
Three community members at a shared worktable in a public library: a teenager tightening a loose bicycle light, an older adult stitching a canvas tote, and a librarian testing a small table lamp. They exchange advice and tools.
This version makes the number, roles, objects, and relationship observable. The mix of repair tasks broadens the story. “Exchange advice and tools” supports cooperation rather than three isolated portraits.
We should still verify whether the actions and tools are plausible. If accurate repair details are central, a workshop facilitator should review them.
Read
Step 2: direct composition
The website places headline copy on the left and registration controls below it. The people should occupy the right half:
Wide 16:9 frame at eye level, medium-wide view. Grouped activity fills the right two-thirds; the left third is a simple, low-contrast wall and softly blurred shelves reserved for headline and button. All six hands and the three repair objects remain visible. The librarian’s welcoming expression and the shared tools form the focal path.
“All six hands” may still fail—counting and anatomy are unstable—but it becomes a testable requirement. The copy zone is described positively, not only as “no clutter.”
Read
Step 3: establish style
The library cannot obtain model releases for realistic synthetic people and does not want viewers to mistake the scene for documentary photography. It chooses an illustrative treatment:
Contemporary editorial illustration with simplified geometric figures, medium-weight ink outlines, subtle recycled-paper grain, warm cream background, deep navy shadows, and coral and sage accents. Soft evening window light; optimistic and practical, not childish or glossy.
These properties create a reusable visual system. The description avoids naming an artist. The illustrative medium also reduces—but does not eliminate—deception and identity concerns.
Read
Step 4: add hard constraints
The final must-not block targets observed and foreseeable failures:
No generated words, letters, numbers, logos, brand marks, wall posters, or product packaging. No extra people, duplicated tools, fused hands, cropped fingers, unsafe exposed wiring, or sharp tools pointed toward another person. Keep the entire left copy zone free of faces and high-contrast objects.
Some exclusions need positive support. The prompt already describes the left area and the visible activities. Typography and the library logo will be added later in the approved design system.
Read
The complete prompt
The production prompt now reads:
Three community members at a shared worktable in a public library: a teenager tightening a loose bicycle light, an older adult stitching a canvas tote, and a librarian testing a small table lamp. They exchange advice and tools.
Wide 16:9 frame at eye level, medium-wide view. Grouped activity fills the right two-thirds; the left third is a simple, low-contrast wall and softly blurred shelves reserved for headline and button. All six hands and the three repair objects remain visible. The librarian’s welcoming expression and shared tools form the focal path.
Contemporary editorial illustration with simplified geometric figures, medium-weight ink outlines, subtle recycled-paper grain, warm cream background, deep navy shadows, and coral and sage accents. Soft evening window light; optimistic and practical, not childish or glossy.
No generated words, letters, numbers, logos, brand marks, wall posters, or product packaging. No extra people, duplicated tools, fused hands, cropped fingers, unsafe exposed wiring, or sharp tools pointed toward another person. Keep the entire left copy zone free of faces and high-contrast objects.
Longer is justified here because every block has a production role. Remove a sentence only if its requirement is unimportant or handled elsewhere.
Read
Evaluate the revised batch
Use gates before preferences:
- Correct wide composition and usable copy zone.
- Exactly three visible participants and three distinct repair activities.
- No text-like marks or brand resemblance.
- No obvious anatomy, tool, electrical, or safety defect.
- Style aligns with the approved illustration system.
- Scene feels collaborative and welcoming.
Suppose candidate B passes everything except the lamp cord becomes tangled with the sewing thread. That is a local defect. Masking and regional regeneration may be better than rewriting the prompt. After repair, recheck hands, shadows, cord continuity, and the copy zone.
Candidate D may look more dramatic but place a bright lamp on the left. It fails the layout gate even if stakeholders like it. Candidate E may have perfect composition but show four people. The rubric keeps preference from silently replacing the brief.