Describing an image that does not exist

The hardest part of image generation is not the tool. It is that you have to describe something you have only seen in your head, to a system with no way to ask what you meant.

The good news is that five slots cover most of it.

The five slots

Subject. What is in the picture, specifically. Not a dog — an elderly greyhound, curled, asleep.

Setting. Where, and what surrounds it. On a worn leather sofa by a window.

Composition. Where the camera is. Close crop, from slightly above, subject filling two thirds of the frame. This is the slot people skip and the one that most changes the result.

Lighting. Late afternoon light from the left, long shadows. Light does more for mood than any adjective you could reach for.

Medium. 35mm photograph, slight grain. Or watercolour, or pencil study, or 3D render. Each of these carries an entire set of conventions with it.

Assembled: An elderly greyhound curled asleep on a worn leather sofa by a window. Close crop from slightly above. Late afternoon light from the left, long shadows. 35mm photograph, slight grain.

That is a sentence a person could draw from. That is the bar.

Adjectives of quality do nothing

Beautiful, amazing, high quality, masterpiece, stunning, 8K.

These describe your reaction, not anything visible. They pull towards the average of whatever the model associates with praise — which is usually over-saturated, over-lit and slightly plastic.

Replace each with something you could photograph. Not dramatic but single light source, deep shadow on the right. Not high quality but sharp focus on the eyes, background falling off.

This is E3·L3’s rule — context beats cleverness — arriving in visual form.

Medium is the strongest single word

If you change one thing, change the medium.

Watercolour brings soft edges, visible paper, colour bleeding. Blueprint brings line weight, annotation, a limited palette. Polaroid brings a specific colour cast, vignetting and a square frame. One word does the work of a paragraph, because each names a whole tradition with its own conventions.

The corollary: if you do not name a medium, one gets chosen for you — usually a glossy digital-illustration look, because there is an enormous amount of it.

Say it once

Long prompts are not stronger prompts. Repeating an idea in three ways (moody, atmospheric, dramatic lighting) does not triple its weight; it crowds out your other slots.

Say each thing once, clearly, and use the space for a slot you have left empty.

Try it now (7 minutes)

Write a prompt with all five slots filled. Generate it.

Then change only the medium and generate again. The distance between those two images is the single most useful thing to see in this lesson.

Check your understanding

1. Which slot do people most often omit, with the biggest effect? Subject Composition and framing Colour

2. Why do words like 'stunning' and 'high quality' weaken a prompt? They are too long They describe a judgement rather than anything visible Models filter out praise words

3. Naming a medium such as 'watercolour' is powerful because: It is a rare word in training data It carries a whole set of visual conventions at once It reduces the number of generations needed

Recap

Fill five slots — subject, setting, composition, lighting, medium — and say each thing once. Replace judgement words with things you could photograph, and name a medium, because otherwise one is chosen for you. Then change only the medium to see how much a single word carries.