How to write an image prompt that gets what you pictured
The five parts of an image prompt, a weak and a fixed version of each, and the habit that stops a good prompt decaying turn by turn.
Length is rarely what is wrong with a prompt that returns the wrong picture. Usually, the prompt leaves a specific choice open. Five things decide what appears, and any one of them left unsaid gets decided for you: the subject, what it is doing, where it is, what the picture is made to look like, and where the camera stands.
One piece of machinery is worth knowing first, and only one. Your sentence goes to the chat model you selected, which writes an image prompt in its own words and calls a separate tool; that tool carries the prompt and, when you ask for a change, up to three pictures already in the conversation, so the image model never sees the rest of the conversation or the words you typed. On the website Intent Detection is on by default, so the app reads your message and works out that an image is wanted with no button pressed first; the image button and its configuration popover return if you switch that off in Settings, then Tool Preferences. The mobile apps always show a Create image control, which needs a paid plan. A request produces one image.
The sentence is almost all of what you control. Each of the six entries that follow gives a weak prompt, the same prompt fixed, what changed, and a verdict, with the makers' own published advice quoted where it applies.
The five parts, and the habit that saves the most turns
Subject: get specific about the object
Weak: a coffee mug
Fixed: a matte black ceramic coffee mug with a thick handle and a slightly uneven rim, the kind a small studio would throw by hand
What changed: material, finish, and one detail that gives the object a character. Google's image generation documentation puts this first among its best practices: "Be hyper-specific: The more detail you provide, the more control you have." Its own example swaps "fantasy armor" for a clause naming the etching, the collar and the shape of the pauldrons. The same page adds that you should say what the image is for: "Create a logo for a high-end, minimalist skincare brand" will yield better results than just "Create a logo".
Verdict: if two very different objects both fit your sentence, the chat model writing the prompt picks one of them without asking.
Action: give the subject something to do
Weak: a woman with a bicycle
Fixed: a woman in her sixties pushing a steel touring bicycle up a cobbled hill, both hands on the handlebars, looking ahead rather than at the camera
What changed: a verb, plus gaze and hands. OpenAI's image prompting guide asks for those specifics by name, telling you to describe "body framing, relative scale, gaze, and interaction with objects", and its worked examples are instructions of the form "full body visible, feet included," or "hands naturally gripping the handlebars."
Verdict: a noun with no verb returns a catalogue pose, centred and facing front.
Setting: a room with a layout, light with a direction
Weak: in a nice kitchen, cosy atmosphere
Fixed: in a narrow galley kitchen with white tiles and one window on the left, late afternoon sun coming in low and warm, a single lamp on above the sink
What changed: the room has a shape and the light has a source, a time of day and a temperature. "Cosy" describes how you hope to feel about the result, which is not an instruction anything can act on. OpenAI's guide names the substitution to make: for wide, cinematic, low-light, rainy or neon scenes, "specify scale, atmosphere, and color instead of relying on mood words alone".
Verdict: light direction moves a picture further than any adjective you could put in its place.
Style: name a medium
Weak: professional, high quality, 4k, trending
Fixed: a watercolour illustration on rough paper, visible brush edges, a limited palette of ochre, slate blue and off-white
What changed: the words now name something that exists and can be imitated. Google's page on illustration work tells you to "Describe the artistic style, subject, and medium", and its template then asks for the "visual qualities, e.g., bold outlines, cel-shading". If a photograph is what you are after, say it outright: OpenAI's guide tells you to request "photorealistic" or "real photograph" explicitly when that is the goal, rather than assuming it.
Verdict: quality-inflation words name no medium, so they read as decoration and get dropped.
Framing: words for the composition, the control for the shape
Weak: make it wide, close up on her face
Fixed: a medium close-up at eye level, head and shoulders filling the frame, shallow depth of field so the background falls away
What changed: one framing instruction instead of two that fight each other, and the shape moved out of the sentence. Writing "make it wide" leaves the frame exactly where it was. Shape lives in the image configuration control, which is behind the same Intent Detection switch as the image button, and what it offers depends on the maker: the Google and xAI models take ratios, at 1:1, 3:4, 4:3, 9:16 and 16:9, while the OpenAI models are labelled Square, Portrait and Landscape and resolve to fixed pixel sizes. The OpenAI models are also the only place a quality setting appears, at low, medium or high. Beyond the model list, those are the only dials, which is why the sentence carries so much. Camera language still earns its place inside it, with one caveat from OpenAI's guide: treat camera specifications "as cues for appearance, not a guarantee of exact physical simulation".
Verdict: composition is a sentence, shape is a setting, and confusing the two is how a subject ends up letterboxed inside a square.
Iterating: restate the picture instead of appending to it
Weak: same but with a red door
Fixed: the whole prompt again, in one go, with "a red door" written into it
What changed: the details you liked are still in the prompt. Since 12 October 2026 a reply like the weak one can start from the previous picture, because the chat model can hand that image to the image model to edit. The prompt that goes with it is still written from a short description of the earlier attempt, though, so whatever you do not restate can drift. Google's own advice is to iterate conversationally, and for consistency to "include previously generated images in subsequent prompts". An edit now covers the image half of that advice, and restating the working version covers the prompt half. OpenAI's guide supplies the discipline: refine one thing at a time and inspect the result.
Verdict: keep the working prompt in a note outside the chat and edit that text instead of the conversation.
Run one prompt across three makers before you form an opinion
We run image models from OpenAI, Google and xAI, so one experiment is worth the turns it costs: send the prompt you have just fixed to one model from each maker, unchanged, and read the three results together.
Read for the shape of the difference:
how much of your wording survives literally, and how much arrives interpreted;
what each one reaches for in the parts you left unsaid;
default colour, contrast, and how crowded the frame comes back;
how small detail and lettering hold up at the size you actually need;
whether the shape vocabulary fits the job, since a ratio and a fixed pixel size promise different things.
We will not name a winner, and a ranking offered without the prompt attached is worth little. Claims about which model handles faces, lettering or photorealism best move with every release, and the answer that matters is the one your own subject and wording produce.
One thing to plan around: on the free plan you get two image generations every 30 days, they run on Google Nano Banana whatever the picker shows, the picker's ratio and quality choices do not apply on that path, and image generation there is web only. A three-way comparison needs a paid plan, priced in the plans section of the homepage.
Verdict: one prompt, three makers, your own eyes, and the question stops being interesting.
Four fights worth giving up
Exact text inside the image
Lettering is the most common thing to come back subtly wrong, and both makers publish a workaround rather than a fix. Google's documentation files it under limitations: "When generating text for an image, Gemini works best if you first generate the text and then ask for an image with the text." OpenAI's guide says to put required wording in quotes, describe its position and typography, spell unusual words or brand names letter by letter when needed, and then check spelling and legibility in what comes back.
Verdict: winnable on short wording only, and only if you settle it in the conversation first, quote it in the prompt, and read the result closely.
Telling it what to leave out
A prompt that says "no cars" has put cars into the sentence. Google names the fix and gives it a label, recommending "semantic negative prompts": "Instead of saying 'no cars,' describe the intended scene positively: 'an empty, deserted street with no signs of traffic.'"
Verdict: a positive description of the state you want does the work an exclusion cannot. Empty, bare, unbranded, single colour, swept. Keep the explicit ones for what a positive description cannot reach, such as no lettering anywhere in the frame.
The same face twice
Since 12 October 2026 you can ask for a change to an earlier picture, or attach a photo, and the chat model hands it to the image model as a reference: up to three images at a time, or one on Grok Imagine. Google's documentation confirms what makes consistency work where it does work, which is feeding earlier images into later prompts. The chat model decides when to send a reference, so name the picture you mean, for example "keep the woman from the first image, now standing at the door".
Verdict: winnable now, with a reference. Point at the earlier image each time, and still carry the same three or four clauses describing the character into every prompt in the series, because the prompt is written fresh on each turn.
Anything that has to be measured
Camera specifications, floor plans, charts and labelled diagrams all come back plausible well before they come back correct. OpenAI's guide flags both halves of this, treating camera terms as appearance cues and telling you, for process diagrams and information graphics, to "verify labels and factual relationships as well as appearance".
Verdict: treat a generated diagram as a layout draft and check every label yourself. Where a dimension matters, put the number in the sentence, then confirm it in the picture.
The short version
Write the five parts once, as one paragraph, and keep it. When the result misses, change one part of that paragraph and send the whole paragraph again. Most of the frustration people describe with image generation comes from a prompt that was right three turns ago and has been shedding a clause per turn since. Keep the full brief in each request.