How to Write a Better AI Image Prompt (With Examples)

A good AI image prompt describes the picture you want in a fixed order: the subject, then the setting, then the light, then the style, then the camera. Vague results almost always come from leaving out one of those, and the model filling the gap with something you did not choose.

Disclosure: VdoBloom is our platform.

The five parts

1. Subject: who or what is in the picture, and what they are doing. “A woman in a denim jacket”.
2. Setting: where it is. “standing on a rooftop at dusk, city towers behind”.
3. Light: the lighting, which changes everything. “warm low sunset light from the left”.
4. Style: the look. “editorial photograph”, “oil painting”, “anime cel shading”.
5. Camera (optional): framing and lens. “close-up, shallow depth of field” or “wide shot, 35mm”.

Writing them in that order is not a magic formula; it is a checklist. Each part closes off a decision you would otherwise be handing to the model. If you can read your prompt and point to all five, you have told the model almost everything that matters.

Before and after

• Weak: “a cat”. The model picks the breed, colour, background and style for you.
• Better: “a ginger tabby cat sitting on a windowsill in warm morning light, shallow depth of field, editorial photograph”. Now almost every choice is yours.

The difference between the two is not length — it is that the second one answers the five questions the first one left open.

Light: the part most people skip

If you add nothing else to your prompts, add light. It is the fastest way to move from flat and generic to something that looks considered. Useful phrases to keep in hand:

• Direction: “from the left”, “from behind”, “overhead”, “backlit”.
• Quality: “soft and diffused” versus “hard and directional”.
• Time of day: “golden hour”, “blue hour”, “harsh midday”, “overcast”.
• Source: “window light”, “neon signage”, “warm lamp”, “studio softbox”.

“Studio lighting” alone is weak because it says nothing about direction. “Studio lighting, key from top-left, soft shadow to the right” gives the model a scene it can build.

Style and camera: useful vocabulary

Style and framing are where you choose the medium and the crop, and both are easy to name once you have the words.

• Style: photograph, illustration, oil painting, watercolour, 3D render, anime, comic, flat vector, film still.
• Framing: close-up, medium shot, wide shot, full body, overhead, low angle.
• Lens feel: shallow depth of field, wide-angle, telephoto, macro. These change how the scene reads, not just how it is cropped.

Mistakes that cause vague results

• Leaving out the light. Lighting is the single biggest cause of a flat, generic image — always say it.
• Stacking contradictions. “realistic and cartoon and watercolour” gives a muddle. Pick one style.
• Counting on the model for text. Do not put words you need spelled into a prompt; add them afterward (see why AI can't spell).
• Rewording between runs. If you are iterating on one result, keep the prompt identical and change one thing at a time, so you know what caused the change.
• Naming what you do not want. Describing an absence is weak; instead, positively describe the thing you want in its place.
• Front-loading adjectives. A pile of unconnected adjectives before the subject reads as noise. Say what the picture is first, then refine.

Three prompt patterns you can copy

Once you have the five parts, the fastest way to improve is to start from a pattern and fill in the blanks. Three that cover most everyday jobs:

1. The portrait. “[Subject and action], [setting], [directional light], [style], [framing].” Example: “a baker dusted with flour, looking up from a workbench, warm window light from the right, documentary photograph, medium close-up”.
2. The product shot. “[Product] on [surface], [light], [background], [style].” Example: “a glass bottle of olive oil on a stone counter, soft side light, plain warm wall behind, editorial product photograph”.
3. The scene. “[Wide subject], [time of day], [atmosphere], [style], [framing].” Example: “a fishing boat returning to a small harbour, blue hour, thin mist over the water, cinematic photograph, wide shot”.

Notice each pattern still answers all five parts. The pattern just removes the blank page.

Working from a reference or an existing image

If you already have an image you like, you are not limited to words. Reference-capable tools let the model copy structure from an input while you describe the change. When you use one:

• Say what to keep and what to change. “Keep the pose and framing, change the background to a city street at night” is clearer than a general rewrite.
• Still name the light. A reference helps with shape and composition, but direction and quality of light still need to be said.
• Change one thing at a time when you are refining, exactly as you would with a pure text prompt.

When to change the model instead of the prompt

A good prompt cannot rescue a model that draws the thing badly. If you have written all five parts clearly and the output still misses — a style it cannot render, hands that keep failing, detail it flattens — the prompt is no longer the problem. Try the same prompt on a different model before you rewrite it. Comparing two models on one prompt tells you which half of the problem is the wording and which half is the engine. Only change both at once if you are starting over.

A repeatable workflow

1. Write the five parts in order.
2. Generate once. Read the result against each part: is the light there? the right style?
3. Change one part and generate again. Repeating with a single change is how you steer toward what you want.

When a run is close but not right, resist the urge to rewrite the whole prompt. Keep the sentence that worked, change the one clause that missed, and compare the two results side by side. That way the prompt improves with every run instead of resetting to random.

Frequently Asked Questions

Do longer prompts work better?

Specific beats long. A tight sentence that names the light and style beats a paragraph of adjectives.

Should I mention the camera?

Only if the framing matters. “Close-up” or “wide shot” is often enough.

Why do my images all look the same?

Usually because you leave the light and style unspecified, so the model defaults to the same safe look.

What order should the parts go in?

Subject, setting, light, style, camera. The order is a habit that keeps you from forgetting a part — the model does not require it, but a consistent order makes your own edits easier to reason about.

Should I describe what to avoid?

Prefer saying what you want. “Plain empty wall behind her” steers better than “no clutter”, because the model builds from the words you give it.

How many times should I rerun a prompt?

Whenever the result is close, keep the prompt and vary one clause. Rerun the identical prompt a couple of times only when you want to pick the best of several takes of the same idea.

Questions people ask about this

Can I just ask the AI to write my image prompt for me?

You can, but you still have to check the result against the five parts. A generated prompt is only useful if it names the subject, setting, light, style and camera — if it leaves the light out, you are back to the model choosing for you. Treat it as a first draft and fill the gaps yourself.

Why do I still get bad results when I describe exactly what I want?

Two common reasons: the light is missing, or the prompt stacks contradictions like “realistic and cartoon and watercolour”. Also check that you are not naming what you do not want — describing an absence is weak, so positively describe the thing you want in its place. If all five parts are clear and it still misses, the prompt is no longer the problem.

How do I add more detail without making the prompt too long?

Specific beats long. Add detail by closing off a decision — a direction of light, a time of day, a framing — rather than piling on unconnected adjectives before the subject. A tight sentence that names the light and style beats a paragraph of adjectives.

How do I describe who is in the image and what they are doing?

Say the person or thing and the action in the same breath, the way the patterns in this post do: “a baker dusted with flour, looking up from a workbench”. That single clause gives the model a subject and something happening, instead of a static noun it has to interpret.

How do I make my images look more aesthetic?

Add light first — it is the fastest way to move from flat and generic to something that looks considered. Then name the style and the framing, since those choose the medium and the crop. If your images all look the same, it is usually because the light and style are unspecified and the model defaults to the same safe look.

What is a simple process I can follow every time?

Write the five parts in order, generate once, then read the result against each part: is the light there, is the style right? Change one part and generate again, keeping the sentence that worked and changing only the clause that missed. That way the prompt improves with every run instead of resetting to random.

→ Try a prompt · Create a free account