How to Write Great Prompts for AI Image Generators

The gap between what most people get from AI image generators and what is actually possible comes down almost entirely to one thing: the prompt.

Most people type something like “a dog on a beach” and get a dog on a beach. It is technically correct and visually mediocre. It looks exactly like a default AI image, soft, generic, and interchangeable with a thousand others generated from the same three words. Then they see someone else generate a cinematic, breathtaking image from the same tool and assume the software is doing something different.

It is not. The software is the same. The prompt is different.

Writing a good prompt for an AI image generator is a learnable skill, and it improves fast. The principles are not complicated once you understand what these models actually respond to. Here is everything you need to know to turn generic outputs into images that look exactly like what you pictured.

Understand What You Are Actually Doing

Before getting into technique, the most useful thing to understand is what happens when you submit a prompt.

AI image generators do not search a database for an image that matches your description. They generate a completely new image by gradually refining a field of random noise into a coherent picture, guided by your prompt. The model learned from billions of image-caption pairs which visual concepts correspond to which words. When you type “sunset,” the model knows what sunsets look like. When you add “golden hour,” it knows what that specific quality of late-afternoon light looks like. Each detail you add steers the generation toward a more specific visual region.

The practical implication: more specific details generally produce better images, but only if those details are the right ones. Typing a thousand vague adjectives (“beautiful, stunning, amazing, incredible”) does almost nothing because those words do not describe visual properties. Typing five precise, visual descriptors produces a dramatically different and better result.

The other thing to understand is that different tools have different preferred prompt styles. Midjourney V7 responds best to short, high-signal phrase sequences: dense, keyword-rich, without long sentences. ChatGPT’s image generation (GPT-4o) works best with natural language paragraphs. Stable Diffusion 3.5 rewards structured, weighted keywords. Ideogram is the best tool for generating images that contain readable text. Knowing which tool you are using and adapting your style accordingly is as important as the prompt itself.

Start With a Clear Subject and Setting

Every good prompt starts with a clear answer to two questions: what is the main subject, and where is it?

“A woman” is not enough. “A woman in her 40s with silver hair, wearing a tailored navy suit, standing in a rain-soaked Tokyo street at night” is a subject and setting that the model can work with. Each detail eliminates ambiguity and steers the output toward something specific.

The subject-setting combination is your foundation. Everything else you add will modify it, so it needs to be solid first. If you are vague here, all the lighting and style descriptors in the world will not rescue the prompt.

A practical test: can you picture the image clearly in your mind before you type it? If you cannot, the model cannot either. Spend a moment forming a clear mental image before writing, then describe what you see.

Specify the Visual Medium and Style

One of the highest-impact additions to any prompt is specifying the kind of image you want. This is not about the subject. It is about the image’s visual language.

Photography, oil painting, watercolour, pencil sketch, digital illustration, concept art, cinema still, product photography, architectural visualisation: each of these activates a completely different visual vocabulary in the model. Without this direction, the model picks a default, and defaults are rarely what you actually wanted.

For photographic styles, the most effective terms reference specific photographic qualities:

Camera settings: “shot on a 50mm lens,” “f/1.8 bokeh,” “wide-angle,” “macro close-up,” “telephoto compression”

Film and rendering: “shot on film,” “Kodak Portra 400,” “8K resolution,” “RAW photo,” “hyperrealistic”

Photography genre: “editorial photography,” “documentary style,” “studio product photography,” “fashion photography,” “environmental portrait”

For artistic styles, you can reference art movements (“Impressionist,” “Art Deco,” “brutalist,” “ukiyo-e woodblock print”), specific media (“oil on canvas,” “watercolor on paper,” “charcoal sketch,” “ink illustration”), or the general quality of the work (“concept art,” “illustration for a children’s book,” “vintage travel poster”).

One technique that works consistently across platforms is to reference a specific visual genre rather than trying to describe every element. “Shot like a 1970s National Geographic photograph” carries more information in seven words than a paragraph of individual descriptors.

Control Lighting Like a Photographer

Lighting is the single most powerful visual variable you can control in a prompt, and it is the most consistently neglected. The same scene with different lighting looks like an entirely different image.

The most useful lighting terms, in roughly increasing order of dramatic effect:

Soft and natural: “diffused natural light,” “overcast sky,” “north-facing window light,” “soft box lighting,” “overcast natural daylight”

Golden and warm: “golden hour,” “sunset light,” “magic hour,” “warm afternoon light,” “low angle sunlight”

Dramatic and directional: “side lighting,” “rim lighting,” “Rembrandt lighting,” “chiaroscuro,” “hard directional light”

Cinematic and moody: “neon light,” “practical lighting,” “candlelight,” “firelight,” “bioluminescent glow”

Technical and commercial: “studio lighting,” “three-point lighting,” “product lighting,” “even exposure”

You do not need to know photographic lighting theory to use these terms. What matters is that you have tried them enough times to know which ones consistently produce results you like. Build a personal vocabulary of lighting terms that reliably work for your use cases and reuse them.

Add Compositional Direction

The model will make a compositional decision if you do not. Those default decisions are often adequate and sometimes terrible. Giving compositional direction puts you in control.

Framing: “close-up portrait,” “waist-up shot,” “full body,” “aerial view,” “bird’s eye view,” “worm’s eye view,” “over the shoulder”

Aspect ratio: Most platforms let you set this as a parameter rather than a prompt term (Midjourney’s --ar flag, for example), but you can also describe it: “cinematic widescreen,” “square format,” “vertical portrait format”

Rule of thirds and composition: “subject in the left third of the frame,” “centred symmetrical composition,” “negative space on the right,” “leading lines toward the horizon”

Depth: “shallow depth of field,” “deep focus,” “sharp foreground, soft background”, “bokeh background”

Do not try to specify composition and framing and depth of field all at once in your first prompt. Pick the one that matters most for the image you are trying to create. You can always iterate.

Define the Mood and Colour Palette

Two images of the same subject can feel completely different depending on their respective moods and colours. A portrait in warm amber tones with soft lighting reads as nostalgic and intimate. The same portrait in cool blue tones with harsh shadows reads as tense and cinematic. Mood and palette are separate from subject and setting, and they dramatically change the emotional impact of the result.

For mood: “melancholic,” “joyful,” “tense,” “serene,” “mysterious,” “whimsical,” “gritty,” “ethereal,” “nostalgic,” “futuristic”

For the colour palette, be  specific about the general: “warm amber and deep brown tones” produces a more consistent result than “warm colours.” “Desaturated with teal and orange accents” is a recognisable cinematic colour grade that the model knows well. “Pastel mint and blush pink” is specific enough to be useful.

If you have a reference image whose colour palette you want to match, many platforms allow you to upload it alongside your text prompt. This is more reliable than trying to describe a specific palette in words.

Use Negative Prompts to Remove What You Do Not Want

Many platforms, particularly Stable Diffusion and its derivatives, support negative prompts: a separate field for describing what you do not want in the image.

Negative prompts are among the most effective tools for improving output quality, yet beginners underuse them. The key is to be specific. “Bad quality” tells the model almost nothing it can act on. “Motion blur, chromatic aberration, overexposed highlights, watermark, text overlay, extra fingers, deformed hands” targets concrete visual problems the model knows how to avoid.

Common things worth putting in negative prompts for photorealistic images: “cartoon, illustration, painting, watermark, text, logo, blurry, out of focus, overexposed, underexposed, grainy, noise, artefacts, deformed, extra limbs, mutated hands”

For artistic images: “photorealistic, photograph, 3D render” (if you want a clearly artistic style rather than hyper-real)

The most effective negative prompts are built from observation. When you get a result with a specific problem, add the description of that problem to your negative prompt and run again.

Adapt Your Style to the Platform

This is the point most prompt guides skip, and it matters significantly.

Midjourney V7 responds best to short, dense phrases separated by commas, not full sentences. Keep each descriptor to two to four words. Reference images attached to the prompt are the most reliable way to anchor a specific aesthetic. Parameters like --ar (aspect ratio), --v (version), and --style fine-tune behaviour outside the main prompt.

Example approach: “Editorial product shot, luxury skincare, marble surface, golden hour, soft bokeh, cream white palette, high fashion magazine --ar 4:5 --v 7”

DALL-E (via ChatGPT) works best with descriptive paragraphs written in natural English. It also supports multi-turn editing: you can generate an image, then ask it to change specific elements in follow-up messages without having to start from scratch. This iterative conversation approach is one of its strongest features.

Example approach: “Generate a photograph of a 1970s American diner interior at night, shot from a booth looking toward the counter. The lighting should be warm fluorescent with neon signs visible through rain-streaked windows. Film grain, slightly desaturated, cinematic aspect ratio.”

Stable Diffusion rewards structured keywords and benefits most from negative prompts. The CFG scale (classifier-free guidance scale) controls how strictly the model follows your prompt. Higher values produce more literal adherence; lower values give the model more creative freedom.

Ideogram is the tool of choice when your image needs to contain readable, well-formed text. Other generators consistently struggle with typography. Ideogram was built specifically to handle text rendering, and it does it dramatically better than the alternatives.

Iterate, Do Not Restart

The most common mistake beginners make is starting from scratch every time they get an unsatisfying result. That approach throws away useful information.

When an image is mostly right but wrong in one specific way, change only that one thing and run it again. If the composition is good but the lighting is wrong, adjust only the lighting descriptor. If the style is right but the colour palette is wrong, change only the palette description. This systematic iteration is how you learn what works and move toward the image you want.

Save prompts that produce good results. Most platforms store your generation history, but maintaining your own document of prompts that worked, along with what they produced, builds a personal library you can reference and remix.

When a seed number is available (e.g., Stable Diffusion or some Midjourney workflows), save it. Seeds allow you to reproduce a specific composition while varying other elements, keeping the underlying structure the same. This is how you maintain consistency across a set of images.

A Simple Framework to Remember

If none of the above has yet clicked into a usable structure, here is one that works across every major platform. Think of a good prompt as having six components, and ask yourself whether you have addressed each one:

Subject and setting: Who or what is in the image, and where?

Medium and style: What kind of image is it (photograph, painting, illustration)?

Lighting: What is the quality and direction of light?

Composition: How is the frame organised?

Mood and palette: What does it feel like, and what colours dominate?

Quality descriptors: What level of detail and resolution are you aiming for?

You do not need all six in every prompt. A simple subject and a strong style choice will outperform a vague prompt with all six components filled in. But when you are not getting what you want, this framework is the diagnostic tool: which of these six things did you not specify, and what does the model default to when you leave it out?

The first prompt is rarely the best. The skill is in knowing how to improve it.