Guide · AI images
How to Write Prompts for AI Image Generators That Actually Work
An AI image prompt is a short written brief that tells an image generator what to draw, where, and in what style. The prompts that work best name one clear subject, put it in a setting, pick a style, and then get refined over several rounds instead of rewritten from scratch. This guide gives you a four-part formula, the camera and lighting vocabulary that steers photorealistic results, rules for putting text inside images, and a checklist for diagnosing a bad result. It applies to most current generators, because the major vendors describe very similar habits in their own documentation.

Key takeaways
- Start every prompt with the subject, then add the setting and the style; this order matches how Google describes the basics in its image prompt guide.
- Treat the first image as a draft: change one thing per round rather than rewriting the whole prompt.
- Photographic terms such as lens length, lighting and camera angle push results toward realism, and art terms such as medium and line weight push them toward illustration.
- Keep any text inside an image to a few short words, and expect to regenerate several times.
- Prompts are instructions, not guarantees: the same words can give different images, so save the prompts and settings that worked.
What is a good AI image prompt made of?
A good prompt is a subject, a setting and a style, written as plain description. Google's image prompt guide for Vertex AI gives the same three starting points: the subject (an object, person, animal or scene), the context or background it sits in, and the style, which can be general such as photograph or sketch, or specific such as pastel painting or isometric 3D.
Everything else is a modifier on top of that base. Modifiers cover framing, lighting, lens, film look and quality terms. You add them only when the base result is close but not right.
Write the way you would brief a photographer or illustrator who has never met you. Name what is in the frame, what surrounds it, and what the finished piece should feel like. Short, concrete nouns beat long strings of adjectives.
What formula should I use for the first draft?
Use one sentence in the order subject, setting, style, light. For example: a ceramic coffee cup on a wooden desk by a window, product photograph, soft morning light. That sentence has everything a generator needs and nothing it has to guess at.
Google's Gemini image documentation offers a template for photorealistic scenes: a photorealistic [type of shot] of a [subject description] in a [setting description]. It then says to describe the light and to choose a camera angle and lens. The same page gives a separate rule for illustrations and stickers: describe the artistic style, subject and medium, and be specific about visual detail such as bold lines and colors for consistent results.
Here is the formula as a fill-in list you can copy into a notes app.
- Subject: one main thing, with its most important attribute (a red enamel kettle, not just a kettle).
- Setting: where it is and what is around it (on a stone counter, steam rising, plain wall behind).
- Style or medium: photograph, flat vector illustration, watercolor, 3D render, line drawing.
- Light and mood: soft window light, hard midday sun, warm evening glow, flat studio light.
- Framing: close-up, wide shot, top-down, eye level.
- Format: aspect ratio or intended use, such as a 16:9 header or a square social post.
Not every prompt needs all six lines. Use the first three for a draft and add the rest only to fix a specific problem.
Which camera, lighting and style words change the result?
Photography terms are the most reliable steering wheel for realistic images. Google's guide lists modifier groups: camera proximity (close-up, zoomed out), camera position (aerial, from below), lighting (natural, dramatic, warm, cold), camera settings (motion blur, soft focus, bokeh), lens types (35mm, 50mm, fisheye, wide angle, macro) and film types (black and white, Polaroid).
The same guide pairs subjects with lens ideas. It suggests prime or zoom lenses at 24 to 35mm for portraits, macro lenses at 60 to 105mm for food, insects and plants, telephoto zoom at 100 to 400mm for sports and wildlife, and wide-angle lenses at 10 to 24mm for landscapes and astronomy. Treat those as starting points to test, not rules the model must obey.
Quality modifiers exist too. Google groups them as general (high-quality, stylized), photo (4K, HDR, studio photo) and art or illustration (by a professional, detailed). They nudge the output, but they carry less information than a concrete lens, light or medium, so use them last.
- Realistic product shot: macro lens, softbox lighting, shallow depth of field, plain background.
- Portrait: 35mm lens, window light, soft focus background.
- Landscape: wide angle, golden hour, long exposure, sharp focus.
- Illustration: flat vector style, bold outlines, limited color palette, no gradients.
- Icon or sticker: simple shapes, thick uniform line, centered, plain background.
How do I get text to appear correctly inside an image?
Keep text short and expect to try again. Google's prompt guide says to limit text to 25 characters or less for optimal generation and to avoid exceeding three phrases for cleaner compositions. It also notes that text rendering is still improving, so several regenerations are normal.
You can name the placement, a general font style and a rough size such as small, medium or large. Do not expect an exact font match. If a client needs precise typography, generate the picture without text and add the words in a layout tool.
OpenAI's image generation documentation makes a related point about complexity. It says higher complexity can help with dense text, complex layouts and relationships between objects, but can increase latency and introduce unwanted changes. For a simple sign or label, a lower setting may be enough, and you should compare results for your own task.
How should I iterate when the first image is close but wrong?
Change one thing per round and say what must stay the same. Google's Vertex guide states that iteration is important: begin with a core idea, then refine with more detail until the image matches what you pictured. Google's Gemini documentation goes further and calls multi-turn conversation the recommended way to iterate on images.
OpenAI's documentation gives the same advice for edits: describe the requested change and the details you want to preserve, then review the result before using it. That second half matters. Without it, a generator may redraw parts you were happy with.
A simple loop keeps you from wandering:
- Generate the first draft from subject, setting and style only.
- Name the single biggest gap between the image and your intent (wrong angle, wrong light, cluttered background).
- Edit the prompt or the follow-up message to fix only that gap.
- Compare with the previous version side by side and keep the better one.
- Stop when the remaining flaws are cheaper to fix in an editor than in another round of prompting.
Why does the generator ignore part of my prompt?
Usually the prompt is too long, contradictory or vague about priority. A generator has to balance every phrase. If you ask for a minimalist flat illustration with photorealistic skin and dramatic film grain, something will lose.
Some tools also rewrite your prompt before generating. OpenAI's documentation notes that with its Responses API the main model may revise your prompt, and the revised text appears in a revised_prompt field. If results surprise you, check whether your tool shows the rewritten version.
Masked edits have a similar limit. OpenAI's documentation warns that because masked edits are prompt-based, the model may not follow the mask's exact shape with complete precision. Plan for a cleanup pass in an editor.
- Pick one style per prompt and remove phrases that fight it.
- Move the most important element to the front of the sentence.
- Replace abstract words (beautiful, epic) with concrete ones (low sun, wet cobblestones).
- Cut anything the image would still be fine without.
- Describe what should be present rather than listing what should be absent.
How do I prompt for different kinds of images?
Match the prompt to the job, because a hero photo, an icon and a mockup need different instructions. A blog hero image needs a clear focal point and empty space for a headline. An icon needs simple geometry and a plain background. A product scene needs believable light and contact shadows.
For illustration and icon work, Google's Gemini documentation recommends being specific about visual details such as bold lines and colors. State the medium (flat vector, line art, isometric 3D) and the palette you want. If you need a clean scalable file afterward, read how to convert an image to vector and compare tools in AI image generators.
For photographic work, think like a photographer: lens, distance, light direction and time of day. For editing an existing photo, browse AI photo editors, which are built around changing a picture you already have rather than inventing a new one.
What do I do with prompts that work?
Save them, with the tool, model version and settings. The same words can produce different results across tools and versions, so a prompt without its context is half a recipe. A small spreadsheet or notes file with the prompt, the tool, the aspect ratio and a thumbnail is enough.
Build a short house style from your best results: a preferred lighting phrase, a standard medium, a color palette in plain words. Reusing that block keeps a series of images looking like one set.
Check the license before you publish. What you may do with the output depends on the tool's terms and on copyright law, which is covered in Can You Use AI-Generated Images Commercially?. Generated images from some tools also carry markers: Google's documentation states that all generated images include a SynthID watermark, so do not assume an output is untraceable.
What does a prompt look like after three rounds of refinement?
A worked example shows how small each change should be. Say you need a header image for an article about remote work. Round one: a person working on a laptop at a kitchen table, photograph, soft morning light.
The result is fine but cluttered. Round two keeps everything and adds one instruction: plain light wall behind, empty space on the right side of the frame. Now the layout has room for a headline.
Round three fixes the mood: keep the composition, make the light warmer, and show the laptop screen turned away from the camera so no readable content appears. Three rounds, one change each, and you can trace which edit caused which effect. If round three had gone wrong, you would know exactly what to undo.
What are the most common prompt mistakes?
Most weak results come from a handful of habits. Packing in many subjects is the first: a generator asked for five things gives each one less attention. Mixing styles is the second, such as asking for a flat icon with photographic lighting.
The third is stacking vague praise words. Terms like stunning, masterpiece or epic describe how you feel, not what is in the frame. Replace each one with something a camera or a pencil could actually record.
The fourth is skipping the use case. An image for a small thumbnail needs a bold, simple subject, while an image for a large banner can hold detail. Say where the image will be used and crop space accordingly.
Prompt checklist: what to check before you hit generate
Run this list in under a minute. It catches most of the problems that make people blame the tool.
- Is there one clear subject?
- Does the prompt say where the subject is?
- Is there one style or medium, not three?
- Is the light described?
- Is any text inside the image under about 25 characters?
- Is the format (aspect ratio, crop space for headlines) stated?
- Is every phrase concrete enough to picture?
- Is the intended use cleared against the tool's license?
If you answer no to more than two, rewrite before generating. Browse AI image generators, AI design tools and AI photo editors to compare what each tool offers for editing and iteration.
Frequently asked questions
How long should an AI image prompt be?
Long enough to name a subject, a setting and a style, and no longer. One or two sentences is a good first draft. Add modifiers such as lens, lighting or medium only when the first result is close but not right.
Very long prompts often contain contradictions that make the generator drop parts of them.
Do I need special keywords like 4K or HDR?
No. Google's prompt guide lists terms such as 4K, HDR and studio photo as photo quality modifiers, but they carry less information than concrete choices about lens, light and setting. Try them last, and keep them only if the result visibly improves.
Why does text in my image come out misspelled?
Text rendering is still improving. Google's guide advises keeping text to 25 characters or less and no more than three phrases, and expects several regenerations. If exact wording matters, generate the image without text and add the words in a design tool.
Can I get the same image twice?
Not reliably. Many tools produce a different image each time, and some rewrite your prompt before generating. Save the prompt, tool, model version and settings of anything you like, and use the tool's variation or edit features to build on a result instead of regenerating from scratch.
Should I describe what I do not want?
Usually it is better to describe what you do want. Mentioning an unwanted object can put it in the picture. Some tools have a dedicated exclusion setting, so check your tool's documentation, and otherwise rewrite the prompt to name the replacement, such as a plain wall instead of a cluttered background.
Is a prompt enough to own the copyright to an image?
In the United States, not by itself. The U.S. Copyright Office concluded in its January 2025 report that, given current generally available technology, prompts alone do not provide sufficient human control to make users the authors of the output. Our guide on AI images and commercial use explains what you can do instead.
Sources and further reading
- Image generation prompt guide (Vertex AI), Google Cloud
- Image generation with Gemini, Google AI Studio
- Image generation guide, OpenAI
- Copyright and Artificial Intelligence, Part 2: Copyrightability, U.S. Copyright Office