Image Generation, in PracticeLesson 20 of 69

Contents

Image Generation, in Practice

Lesson 20 of 69. Sign in to keep your place.

Module 2 · From a request to a specification you can generate against

The one page that ends the revision argument

7 min read

What the document is for

Not bureaucracy. The spec exists so that the sentence “this isn’t what we wanted” has somewhere to be resolved. Without it, every review is a negotiation about remembered intentions. With it, a review is a comparison between an image and a page you both agreed to.

It fits on one side. If it runs longer, it has become a proposal and something else has gone wrong.

The template

Markdown
# Image spec — [client] — [date]

## Deliverables
- Hero: 1200 × 1500 px, JPEG, sRGB
- Crops from the same frame: 1080 × 1080, 1080 × 1920
- Print version: A4 at 300 dpi, CMYK, 3 mm bleed

## Generation plan
- Native render ratio: 896 × 1152 (SDXL bucket), upscaled and cropped
- Crop-safe core: central square; headline sits in the upper third

## The picture
One sentence describing the image. Not a prompt — a sentence a person understands.

## Reference pack
1. light — ref-01.jpg — the hard side light, not the styling
2. palette — ref-02.jpg — the muted greens
3. composition — ref-03.jpg — subject low and left, lots of sky
4. product — ref-04.jpg — our actual bottle, photographed 14 Mar
5. texture — ref-05.jpg — matte, fine grain, not glossy
6. anti-reference — ref-06.jpg — too clean, too corporate

## Constraints
- Brand red #E4002B, exact, on the label only
- Headline zone: upper third must be quiet, contrast ratio at least 4.5:1

## Must not appear
- Legible brand names or logos of any kind
- Text other than type we set ourselves
- Hands

## Photographed, not generated
- The bottle (composited into a generated environment)

## Rounds
Two rounds of revisions included. Structural changes — different composition,
different concept — are a new round.

## Disclosure
Client is informed the environment is AI-generated. Delivery includes a note
for their records.

The four lines that do the work

Most of that template is useful. Four lines are load-bearing.

”Photographed, not generated” sets the expectation that some elements are real before anyone has an opinion about the picture. It is far easier to state on day one than to explain on day six.

”Must not appear” is the list a person checks against the finished file. It is the difference between an embarrassment caught in QA and one caught by the client’s legal team.

The rounds clause distinguishes a tweak from a rebuild. “Can we make it warmer” is a tweak. “Can she be facing the other way” is a regeneration that loses everything downstream of the original frame, and the spec is where that gets priced.

The disclosure line means nobody is surprised. The rights block later covers what you must say and where; the point here is that it is agreed at the start rather than discovered at the end.

Getting it signed without ceremony

You do not need a legal process. An email saying “here’s the spec: confirm and I’ll start” and a reply saying “confirmed” is enough for the purpose, which is a shared record rather than an enforceable instrument.

What matters is the timing: before generating. A spec written after the first review is a defence, and everyone can tell.

Its real limit

A spec cannot make a client know what they want. Some do not until they see something, and insisting on precision before any image exists can stall a project that a quick sketch would unblock.

So use it proportionately. For a small job, three lines in an email (dimensions, what must not appear, how many rounds) carries most of the value. For anything with print, legal exposure or more than a few images, write the page. And if a client cannot answer the questions yet, generate three cheap, deliberately rough concepts to make the conversation concrete, then write the spec from whichever one they point at. The spec is a tool for removing ambiguity, not a gate that ambiguity has to pass first.

Keep it alive

A spec written once and never touched becomes fiction by the third day, because briefs move. The version that stays useful is edited as decisions are made, with a short note of what changed and when.

Markdown
## Changes
- 14 Mar: headline moved to lower third — upper third now free
- 16 Mar: bottle to be photographed, not generated (label unreadable in tests)
- 18 Mar: second size added, 1080 × 1920

Three lines like that resolve more disputes than the entire rest of the document, because the arguments that actually happen are almost never about the original brief. They are about a change somebody remembers differently.

A last, unglamorous point: keep the spec beside the work, in the project folder, not in a chat thread. Chat histories are searched badly and scroll away. A file called brief.md sitting next to the images is still there in a year, when someone asks why the composition has all that empty space at the top and nobody remembers that a headline was supposed to go there.

Every lesson here is open. Your progress is recorded once you pass the Vectors, Logos and Things That Get Printed exam.

Check your understanding

What does an image spec actually accomplish at review time?

Pick the one you would defend. Your answer stays private.

Next lesson · 10 minWhat you can sell, and what you must disclose

Stuck on something?Ask Neema, your AI tutor

How to cite: Addaly (2026) ‘The one page that ends the revision argument’, in Image Generation, in Practice. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-images/writing-the-spec (Accessed: 13 October 2026).