How to write AI image prompts that work
A good AI image prompt describes the picture in plain sentences: the subject, what they are doing, where, the light, the camera framing and the style, plus anything that must stay the same if you are editing a photo. Modern image models understand full sentences far better than lists of keywords, and Google's own guidance for its Gemini image models puts it simply: the more specific you are, the more control you have. Start with one clear sentence about the subject, then add detail one layer at a time.
The six parts of a strong prompt
Most good prompts answer the questions a photographer or illustrator would ask before starting. You rarely need all six, but checking them is the fastest way to find what a prompt is missing.
- Subject: who or what the picture is about, with the details that matter ('a woman in her thirties with curly red hair', 'a ceramic coffee cup with a chipped rim').
- Action or pose: what the subject is doing ('laughing while looking over her shoulder').
- Setting: where and when it happens ('a narrow street in Kyoto after rain, early evening').
- Light: the biggest single influence on mood ('soft window light', 'golden hour', 'neon signs reflecting on wet pavement').
- Camera and framing: how close and from where ('close-up portrait, 85mm lens, shallow depth of field', 'overhead flat lay').
- Style: photograph, film still, illustration, 3D render or watercolour, plus any colour palette.
Write sentences, not keyword lists
Older image generators rewarded long strings of tags like 'portrait, 8k, ultra detailed, masterpiece'. Today's models are built on language models and read a prompt the way a person reads a brief. Filler words add nothing, while a sentence about how things relate ('the dog sits at her feet, looking up at the ice cream') says something a tag list cannot.
Compare two prompts. Keyword style: 'woman, cafe, Paris, morning, coffee, cinematic, bokeh'. Sentence style: 'A woman in a camel coat reads a paperback at a small marble table outside a Paris cafe in the early morning. Steam rises from her coffee. Shot on a 50mm lens with the street softly blurred behind her, natural light, warm muted colours.' The second leaves far less to chance, and it is still only three sentences.
Prompts for editing your own photo
When you add a photo, the prompt has two jobs: say what to change and say what to keep. Models follow the change you describe, and anything you leave unsaid is open to reinterpretation.
- Name what stays: 'Keep my face, hairstyle and expression exactly as in the photo.'
- Describe the change, not the whole picture again: 'Change the background to a sunny beach and match the light on my face to it.'
- Make one change per round when refining, so you can see what worked.
- With several reference photos, refer to them by order: 'the person from image 1 wearing the jacket from image 2'.
- Keep the framing close to your photo. Asking for a full-body shot from a tight selfie forces the model to invent your body.
Words that control light and camera
A few photographic terms do a lot of work, because the models have learned from enormous numbers of captioned photos that use them. In Luren AI you can type prompts like these in the Image tab, add up to six reference photos and pick the model that suits the job.
- '85mm portrait lens, shallow depth of field' for a flattering close portrait with a blurred background.
- '35mm film, natural grain' for a candid, analogue feel.
- 'Golden hour backlight' for warm, glowing edges; 'overcast daylight' for soft, even skin tones.
- 'Studio softbox lighting on a plain grey backdrop' for clean headshots and product shots.
- 'Low angle' to make a subject look powerful; 'overhead flat lay' for food and products.
- 'Wide establishing shot' when the place matters as much as the person.
Text, aspect ratio and size
If the image needs words on it, put the exact text in quotation marks and describe the lettering: a vintage poster with the headline 'Summer Market' in bold red serif letters. Models such as GPT Image 2 and Nano Banana Pro are strong at readable text; others may misspell longer phrases, so keep the wording short.
Set the shape in the app's aspect ratio setting rather than in the prompt, and choose it for where the image will be used: 9:16 for Stories and Reels, 4:5 or 3:4 for Instagram posts, 16:9 for YouTube thumbnails. The aspect ratio calculator helps when you need exact pixel sizes.
Describe what you want, not what you don't
Many models treat the words in a prompt as things to include, so 'no people' can still put people in the picture. Describe the result positively instead: 'an empty beach at dawn' rather than 'a beach with no people'. Google gives the same advice in its prompting guide for its Veo video model. If something unwanted keeps appearing, describe what should be in its place.
Iterate like an editor
Treat the first image as a draft. Look at it, decide the single most important thing to fix, and ask for only that. Save prompts that work and reuse them as templates, swapping the subject or the setting.
If you are stuck for ideas, the prompt library has hundreds of tested examples sorted by model and topic, and the free AI prompt generator builds a structured prompt from a few choices.
