How to Write AI Image Prompts That Actually Work
Knowing how to write AI prompts for images is mostly about removing guesswork. A short prompt like a dog in a park leaves the model to decide the breed, the light, the angle and the mood, so every result is a gamble. A structured prompt answers those questions up front. This guide breaks a strong image prompt into six parts: subject, style, lighting, framing and lens, aspect ratio, and what to avoid. Each step explains why it matters and gives an illustrative phrase you can adapt. We use the same structure for every prompt in our library, and when one is tested we publish the real output and the model that made it. The framework works across Gemini, ChatGPT, Grok and Midjourney, and the same thinking carries over to video tools like Sora. Read it once, then use it as a checklist.
The structure of a good AI image prompt
1. Subject
Start with who or what the image is about and what it is doing. The subject carries the most weight in any prompt, so be concrete: age range, clothing, expression, pose and action are all fair game, as long as you describe a fictional person rather than a real, identifiable one. Put the subject in the first sentence so the model treats it as the priority. If there are several subjects, give each one a position and a role, otherwise details tend to swap between them. A clear subject also makes it easy to reuse the rest of the prompt later.
Example phrase · a woman in her thirties in a mustard raincoat, laughing as she holds an umbrella
2. Style
Next, say what kind of image you want. Photorealistic photo, film photograph, watercolor illustration, 3D render and flat vector art are very different targets, and the model cannot read your mind. For photos, naming a photographic genre, such as street photography, editorial portrait or product shot, often works better than adjectives like beautiful or stunning. Avoid naming living artists or copyrighted characters; describe the visual qualities you like instead, such as muted colors, heavy grain or bold outlines. One clear style beats a stack of competing ones.
Example phrase · candid street photography, shot on 35mm film, muted colors, light grain
3. Lighting
Lighting decides the mood and is the fastest way to make an image look professional. Describe the light source, its direction and its quality. Soft window light from the left, harsh noon sun, golden hour backlight, overcast daylight and a single neon sign all produce very different pictures. For realistic images, natural and slightly imperfect light usually looks more believable than perfect studio lighting. If you want drama, mention contrast and shadows directly. If faces look flat, add a light direction and a little rim light to separate the subject from the background.
Example phrase · soft overcast daylight with a warm shop window glowing behind her
4. Framing & lens
Tell the model where the camera is and how close it sits. Shot size, such as close-up, waist-up or wide shot, controls how much of the scene appears. Camera angle, like eye level, low angle or overhead, changes how powerful or intimate the subject feels. For photorealistic prompts, a lens length adds useful cues: around 85mm suggests a flattering portrait with a blurred background, while 24mm feels wide and immersive. Mention depth of field if you want a sharp subject against a soft background, and say where the subject sits in the frame.
Example phrase · waist-up shot at eye level, 50mm lens, shallow depth of field, subject off-center
5. Aspect ratio
Decide the shape of the image before you generate, because it changes the composition, not just the crop. Square 1:1 suits profile pictures and feeds, 4:5 fits tall social posts, 9:16 works for stories and phone wallpapers, and 16:9 suits banners, thumbnails and desktop backgrounds. In chat-based tools like Gemini, ChatGPT and Grok, write the ratio into the prompt in plain words. In Midjourney, use the aspect ratio parameter at the end. If the tool ignores the ratio, ask again in a follow-up or crop the result afterward.
Example phrase · vertical 4:5 composition (in Midjourney: --ar 4:5)
6. Negative prompt (what to avoid)
Finally, name what you do not want, but only when you need to. Negative instructions are most useful for fixing a repeated problem: extra fingers, text or watermarks, a cluttered background or an overly smooth plastic look. In chat-based tools, phrase them as a short plain sentence at the end. In Midjourney, the no parameter handles this. Keep the list short, because a long list of banned items can distract the model from your main idea. Often it is better to describe what you want instead, such as a plain gray background rather than no clutter.
Example phrase · avoid text, logos and watermarks; keep skin texture natural (in Midjourney: --no text, watermark)
Weak prompt vs. strong prompt
Weak
A woman with an umbrella in the city, beautiful, high quality.
Strong
Candid street photograph of a woman in her thirties in a mustard raincoat, laughing as she holds a clear umbrella on a rainy city sidewalk. Soft overcast daylight with a warm shop window glowing behind her. Waist-up shot at eye level, 50mm lens, shallow depth of field, subject slightly off-center. Shot on 35mm film, muted colors, light grain, natural skin texture. Vertical 4:5. Avoid text, logos and watermarks.
How to make AI images look real
- Name a photographic style, such as candid phone photo, documentary photo or film photograph, instead of words like hyper-realistic.
- Add a lens length and depth of field, for example 50mm with a softly blurred background.
- Describe one natural light source and its direction, like window light from the left or late afternoon sun.
- Ask for natural skin texture, pores and small imperfections rather than smooth or flawless skin.
- Use an ordinary, lived-in background with believable everyday details.
- Give the subject a candid action or expression instead of a stiff posed look.
- Remove quality buzzwords like 8K, ultra-detailed and masterpiece, which often push images toward a glossy, artificial look.
- Check hands, text, jewelry and reflections in the result, then fix any problems with one targeted follow-up.
How prompting differs between models
Gemini (Nano Banana)
Nano Banana is the image model inside Google's Gemini app. You write prompts conversationally, so full descriptive sentences work better than keyword lists. It is especially useful for editing an uploaded photo while keeping the subject recognizable. State the aspect ratio in words, and refine results with short follow-up messages that change one thing at a time.
ChatGPT Images
ChatGPT generates images from natural-language requests inside a chat. It follows detailed descriptions and can render short text in images, which helps for posters and labels. You can attach a photo to edit or restyle it. Mention the aspect ratio in your request, and use follow-up messages for targeted changes rather than starting over.
Grok
Grok, from xAI, creates images from prompts typed into its chat on the web, in its app and on X. Plain descriptive sentences work well, and the same subject, style, lighting and framing structure applies. Features such as photo editing and aspect ratio controls can vary by version and platform, so check what your interface offers.
Sora
Sora is OpenAI's video generation model, so prompts need to describe motion as well as appearance. Write one shot at a time: subject, action, camera movement, lighting and sound, plus a sense of pacing such as a 10-second shot. The framing and lighting habits from image prompts still apply, with orientation chosen as portrait or landscape.
Midjourney
Midjourney reads prompts as descriptive phrases and uses parameters at the end for technical settings. Add --ar followed by a ratio, such as --ar 4:5, to set the aspect ratio, and --no followed by items, such as --no text, to exclude things. Keep the description focused on subject, style, lighting and framing, and leave technical settings to parameters.
Want a head start? The free prompt builder assembles these six parts for you, formatted for the model you use.
FAQ
How to prompt AI to generate images?
Describe the image in one or two clear sentences covering the subject, style, lighting, framing and aspect ratio, then paste it into an image tool such as Gemini, ChatGPT, Grok or Midjourney. Review the result and change one detail at a time. Adding what to avoid helps only when the same mistake keeps appearing.
What are some good prompts for AI images?
Good prompts for AI images are specific rather than long. For example, a candid street photo of a laughing woman in a mustard raincoat, soft overcast light, 50mm lens, vertical 4:5. Our AI photo prompts collection has many copy-ready examples for portraits, realistic shots, aesthetic scenes and photo edits, each with a clear structure.
How to make AI images look real prompt?
To make AI images look real, prompt like a photographer. Name a photo style, a lens length, one natural light source and its direction, and ask for natural skin texture and an everyday background. Drop buzzwords like 8K or flawless, which often add a glossy, artificial finish. Then check hands and text closely.
How do I create a prompt?
Create a prompt by answering six questions in order: what is the subject, what style is it, how is it lit, how is it framed, what shape is it, and what should be avoided. Write the answers as short phrases in one paragraph. Our free prompt builder does this for you from simple menu choices.