AI Image Prompts: How to Write Ones That Work (2026)
Learn how to write AI image prompts that work: a simple structure, readable text, edits without drift, and what GPT Image, Nano Banana and FLUX advise.
Good AI image prompts describe what a camera would record: the subject, the setting, the light, the framing and any constraints, in a consistent order. Start with one clear sentence, check the result, then change one thing at a time. Vendor guides all favour specific, visible detail over keyword lists.
How to write an AI image prompt that works
Most weak prompts fail for one of three reasons. They list keywords instead of describing a scene. They ask for things the model cannot see, such as "high quality" or "masterpiece". Or they change several things at once, so you cannot tell which change helped. Fix those three problems and most prompts improve.
AI image prompts: a structure you can reuse
Use this order as a habit. Put the most important information first, because a prompt that starts with the subject gives the model a clear focus.
- Image type and medium: photo, watercolour, 3D render, product mockup, thumbnail
- Subject: who or what is in the frame, and what it is doing
- Setting: where it is, and what is in the background
- Light: the source, direction and quality of the light
- Framing: shot size, camera angle, lens or depth of field
- Constraints: what must stay visible, what must not change, and any text
Black Forest Labs' prompting guide uses almost this order: image type and medium, subject, setting, light, then framing and detail. It adds that "a clear structure matters more than length."
What the big vendors recommend
Each vendor has a slightly different view. Use them as starting points, and test them on your model.
| Vendor | What it recommends | Where it says so | |---|---|---| | OpenAI (GPT Image) | Write in a consistent order: scene, subject, key details, constraints. State the intended use, such as an ad or an infographic. For text, use quotes or ALL CAPS and specify the font, size, colour and placement. For edits, say "change only X" and "keep everything else the same". | OpenAI prompting guide | | Google (Nano Banana) | "Describe the scene, don't just list keywords." Say what the image is for, iterate in small steps, and describe what you want rather than what you don't. | Google Developers Blog, August 2025 | | Black Forest Labs (FLUX) | Use the order above. Avoid generic quality words such as "highly detailed" and tags such as "masterpiece". Start with one sentence and add detail after you see results. | BFL prompting guide |
Google's tips were written in August 2025 for Gemini 2.5 Flash Image, so they do not automatically apply to Nano Banana 2.1 or Nano Banana Pro. Test them.
Before and after: a weak prompt and a strong one
This example is an illustration written for this article. It is not a vendor example.
Weak: "dog, park, nice, 4k, masterpiece, highly detailed"
Stronger: "A golden retriever mid-leap over a puddle in a city park on an overcast morning. Soft, diffuse light. Shot from knee height with a 35mm lens and a shallow background blur."
The second prompt names one subject, one action, one light source and one camera setting. Every detail is something the model can draw. The first prompt gives it a list it has to guess from.
How long should a prompt be?
Long enough to name everything that matters, and no longer. Black Forest Labs says "strong prompts are specific, not necessarily long", and suggests removing any sentence that does not describe something visible. Start with a sentence or two, then add detail only where the first result falls short.
Getting readable text in images
Text is where prompts most often fail. OpenAI suggests putting the literal text in quotes or ALL CAPS, naming the font, size, colour and placement, and spelling out tricky words letter by letter. It also recommends medium or high quality for small text. These two illustrative prompts were written for this article:
- Poster headline: a sign reading "OPEN LATE" in bold white capitals, centred above the door.
- Menu: a chalkboard with the words "Cold Brew" and "Oat Latte" in a hand-lettered style, left-aligned.
Google's model card for Nano Banana 2.1 lists weak rendering of small text and long paragraphs as a known limit. Keep the text short, and check it at the size it will be shown.
Editing without drift: change one thing, keep the rest
Edits go wrong when you ask for too much at once. OpenAI's guide suggests "change only X" followed by "keep everything else the same", and restating the preserve list on each round. Iterate with single-change follow-ups rather than one long instruction.
For example: "Change only the sky to a sunset. Keep everything else the same: the dog, the path, and the light on the dog."
Google's docs recommend multi-turn editing for the same reason. Their example is: "Update this infographic to be in Spanish. Do not change any other elements of the image." Mask-based editing is less exact. OpenAI notes it is "entirely prompt-based", so the model may not follow the mask exactly.
Negative prompts: describe what you want instead
A negative prompt says what to leave out, such as "no people" or "no text". Google's guidance is to describe the result instead. For example, write "an empty street" rather than "a street with no cars". Black Forest Labs gives similar advice.
Whether negative wording works depends on the model. Treat it as something to test, not a rule that never works.
Keeping a character or product consistent
Use reference images where the tool supports them. The Gemini API accepts up to 14 reference images per request. Nano Banana 2.1 can keep up to four characters and ten objects consistent. Google's model card still lists imperfect character consistency as a limit.
If the features drift after many edits, Google suggests starting a new conversation and describing the character in detail again. Only use reference images you have the right to use.
Quality and cost settings: draft cheap, finish high
OpenAI suggests starting with low quality for high-volume or latency-sensitive work, then comparing medium or high for small text, dense infographics, close-up portraits and identity-sensitive edits. The API prices below are per image at 1024x1024, with input tokens billed separately.
| Quality | GPT Image 2 | GPT Image 1 Mini | |---|---|---| | Low | $0.006 | $0.005 | | Medium | $0.053 | $0.011 | | High | $0.211 | $0.036 |
The per-image prices for the 2.5 models are not in OpenAI's table, so they are not listed here.
Common mistakes
- Listing keywords instead of describing a scene
- Using words the model cannot draw, such as "masterpiece" or "ultra quality"
- Asking for several changes in one edit
- Writing a long prompt before you have seen a first result
- Assuming a prompt that works on one model works on another
- Expecting exact text from a model without checking it
Prompts and copyright
Prompting affects copyright, not only quality. The US Copyright Office's January 2025 report says that, based on current technology, "prompts do not alone provide sufficient control" over the output. It compares repeatedly revising a prompt to "re-rolling the dice". Our AI image copyright guide explains what that means in practice.
Try these prompts at OndaLabs
The OndaLabs image studio lets you run one prompt on five models: Nano Banana Lite, Nano Banana, Nano Banana Pro, GPT Image and GPT Image Mini. You can choose an aspect ratio and edit with reference images. The credit cost is shown before you generate, so you can draft on a cheaper model and finish on a stronger one. For example, GPT Image Mini costs 120 credits at low quality and 1,500 at high, and Nano Banana costs 2,000 credits at 1K and 6,000 at 4K.
Image generation is included from the Starter plan, at $9 per month, and commercial use is included on Starter and above. Compare plans on the pricing page.
For a comparison of models, see the best AI image generators in 2026. For Google's models in particular, read our Gemini image generation guide.
Frequently asked questions
How long should an AI image prompt be?
Long enough to name the subject, setting, light, framing and any constraints, and no longer. Black Forest Labs says strong prompts are specific rather than long. Start with a sentence or two, then add detail only where the first result falls short.
Do negative prompts still work?
It depends on the model. Google and Black Forest Labs advise describing what you want instead of what you do not want, such as "an empty street" rather than "no cars". Test negative wording on the model you use before you rely on it.
How do I get an AI image generator to spell text correctly?
Put the exact words in quotes or ALL CAPS, name the font, size, colour and placement, and keep the text short. Check the result at the size it will be shown, and add longer text in an editor if the model still gets it wrong.
How do I edit one part of an AI image without changing the rest?
Ask for one change at a time, and state what must stay the same. For example, "change only the sky to a sunset, and keep everything else the same". Repeat the preserve list on each round.
Should I add words like "masterpiece" or "highly detailed"?
Black Forest Labs advises against generic quality words such as "highly detailed" and tags such as "masterpiece". Describe what a camera would record instead, such as the lens, the light and the framing.