Prompting images

The image model reads prose, not keyword lists. It follows a detailed, concrete description closely, places things where you say they are, and renders written text accurately when you quote it. A prompt can be up to 4000 characters (8000 with the pro model); the best results usually come from 60 to 250 words.

The short version #

Describe the finished picture as someone looking at it would, in the present tense: "The image is a ...", not "create ..." or "I want ...".
Open with the medium, the style, the subject and the background in one sentence.
Then go through the picture in a fixed order: setting, subject (pose, face, clothing, what it holds), surroundings, lighting, overall mood.
Say where things are: in the upper left, across the bottom third, behind her on the right.
Put any text that should appear in the image in "double quotes" and say where it sits and what it looks like.
Leave size and shape out of the prompt. Set them with the aspect_ratio, megapixels or width and height parameters.

Structure #

Opening sentence: medium, style, subject, background. Always name the medium. It decides more than any other word.

A wide cinematic photograph of a lighthouse keeper standing on a cliff path, a storm front and slate-grey sea behind him.

Media: photograph, portrait, poster, illustration, infographic, logo, product shot, storyboard panel, interface mock-up, page, card. Styles: realistic, photorealistic, cinematic, editorial, minimalist, flat vector, watercolour, ink, isometric, hand-drawn, 3D-rendered, retro, cel animation.

Setting and background. What surrounds the subject, near to far. Surfaces and materials, weather, time of day.

Subject. For a person: build, posture, where they look, expression, hair, each garment with its colour and material, what they hold. Give age as a life stage (a child, a woman in her thirties, an elderly man). For an object: class, material, colour, condition, and brand only if you really mean one.

Placement. Use position phrases throughout rather than once: in the foreground, in the upper-right corner, centred, to the left of, tucked behind. Eight to a dozen of them in a detailed prompt is normal.

Lighting gets its own sentence. Source, direction, quality, and what it does: Soft overcast daylight comes from a window on the left, leaving the right side of his face in gentle shadow, with a faint rim of light on his shoulder.

One closing sentence about the whole. Balance, palette, mood: The overall composition is calm and symmetrical, in muted greens and warm greys.

Wording that works #

Concrete over abstract. A chipped enamel mug on a worn oak table beats a rustic kitchen scene.
Colours with a modifier: deep navy, pale cream, muted olive, burnt orange.
Materials by name: brushed steel, coarse linen, frosted glass, wet asphalt.
Count things, and write small numbers as words: three paper lanterns, not several lanterns.
Photographic terms are understood: close-up, wide shot, low angle, bird's-eye view, shallow depth of field, bokeh, backlit, macro, telephoto compression, negative space.
Say what you want, not what you don't. An empty street works better than a street with no people. For something that keeps appearing anyway, say it once at the end: No text or logos appear in the image.
Skip quality boosters. Masterpiece, 8K, ultra detailed, award-winning do nothing useful. Spend the words on content.
Do not put the aspect ratio or resolution in the prompt. The parameters control them.

Standard or pro #

Text to image has two models. Standard is fast and follows @references. Pro renders finer detail and the most faithful lettering, and keeps up with long, precise prompts: posters, packaging, signs, menus, covers and pictures with many described parts. It costs several times more, draws from the prompt alone (no @references), and a negative prompt costs nothing extra with it. Everything below applies to both; pro rewards even more detail.

Regions (pro) #

With the pro model you can pin parts of the picture in place with up to 12 boxes. Each box has a kind and its own description:

object or character: what fills the box, e.g. a red lighthouse on a rocky cliff.
text: the exact words to write, plus how they look, e.g. the words GRAND OPENING in bold white serif capitals.

The prompt still describes the whole picture (style, mood, light, background); the boxes only say what goes where. Over the API, boxes are fractions of the canvas, x, y, width and height from 0 to 1 with the origin at the top left, so they keep their place at any size. Regions cost nothing extra.

Text in images #

Accurate lettering is one of the model's strengths, including multiple lines, signs, posters, labels, charts and interface mock-ups.

Put the exact text in straight double quotes and keep it character for character: a neon sign that reads "OPEN LATE".
For each piece of text give position, look and content: A bold white sans-serif headline across the top reads "THE MIDNIGHT EXPRESS"; beneath it, in small gold capitals, "A journey under the stars".
Describe a line break as a second line, not with a line break inside the quotes.
Write text in the language and script it should appear in. Do not ask for a translation inside the prompt.
Never use placeholders such as some names or lorem ipsum; write the real content, including axis labels and table cells.
Lettering that should stay unreadable: say it is blurred, indistinct or too small to read.
If there should be no text at all, end with The image contains no text.

Negative prompts #

A negative prompt lists what to keep out. With the standard model it switches to a slower guided mode that roughly doubles render time and price (pro always runs guided, so there it is free), and for most pictures a clearer positive description does more. Use one only when a plain prompt keeps producing the same unwanted thing, and keep it to that thing: blurry text, extra fingers.

Images that will become video #

A start frame is the first frame of a shot, so compose it for motion:

Use the aspect ratio the video needs. The start frame sets the frame size of the whole sequence.
Leave room in the frame for the movement you plan: space ahead of a walking figure, sky above something that will rise.
Keep the subject large enough that a face or product stays detailed once it moves.
One clear light source. The video model keeps the lighting it is given.
Avoid small lettering; fine text does not survive motion well.
For 2D styles, name the style precisely and say what it is not: flat hand-drawn cel animation with bold line art and flat colour, no 3D shading. Frame close enough that faces are large. Repeat the same style words in every later prompt of the sequence.

Using references #

To carry a specific character, product or style into an image, add it to the sequence as a reference and mention its @label in the prompt. See Prompting edits and references.

Variations and reproducibility #

Each generation uses a random seed and reports it. Pass the same seed with the same prompt and size to reproduce an image; change only the seed to get a different take on the same prompt. Keep variations in one sequence (pass its sequence_id) so they sit side by side as candidates.

Examples #

Photographic portrait

A medium close-up photograph of a ferry captain in her fifties on the open bridge of a small boat, grey harbour water and a low fog bank behind her. She has short silver hair under a navy wool cap, weathered skin and a calm half-smile, and looks slightly past the camera to the left. She wears a mustard-yellow waterproof jacket with the collar up, beaded with fine rain. Her right hand rests on a brass wheel at the bottom edge of the frame. Soft overcast morning light comes from the right, with a pale rim along her cheek and cap. The background is out of focus with a few blurred mooring lights. The overall composition is quiet and balanced, in cool greys with one warm yellow accent.

Poster with text

A tall Art Deco travel poster for a night train. A silver locomotive curves from the lower right toward the centre through deep blue mountains under a thin crescent moon in the upper left. Across the top, a bold ivory headline in geometric capitals reads "THE MIDNIGHT EXPRESS". Beneath it, in smaller gold capitals, a second line reads "A journey under the stars". Along the bottom edge a narrow gold band carries the small text "Departing nightly at 23:40". Flat colour, clean geometric shapes, generous margins and a clear typographic hierarchy. The overall composition is symmetrical and elegant, in navy, ivory and gold.

Product shot

A square studio photograph of a translucent emerald-green glass perfume bottle standing on pale limestone beside a shallow pool of water. Rippling sunlight passes through the glass and throws green reflections onto the stone to the right. A single olive branch enters from the upper-left corner. The bottle has a square brushed-brass cap and no label. Hard, warm sunlight from the upper left, crisp edges on the bottle, soft shadows. The overall composition is uncluttered with wide empty space on the right. The image contains no text.

Flat illustration for animation

A wide flat hand-drawn cel animation frame with bold black line art and flat colour, no 3D shading, of a small red fox sitting on a mossy log in a birch forest at dawn. The fox is on the left third, facing right, ears up, tail curled around its feet. Pale birch trunks recede into soft mist on the right, and a few orange leaves hang in the air. Warm low light comes from the right edge of the frame, giving the fox a thin golden outline. The overall composition leaves open space on the right for the fox to move into, in soft greens, cream and orange.

Checklist before you send #

Medium and style named in the first sentence?
Each important thing given a position?
Lighting described in its own sentence?
All text quoted exactly?
No size, ratio or quality boosters in the prompt?
Right aspect_ratio and megapixels set as parameters?