How to Write AI Image Prompts

The four-part structure of a strong prompt

Almost every high-quality prompt, across GPT Image 2, Nano Banana, Midjourney and the rest, can be broken into the same four parts. Naming each one explicitly is what separates a usable image from a generic one.

• Subject & action — who or what, and what they are doing ("a chef plating a dish", not just "food").
• Style / medium — photograph, 3D render, watercolor, cinematic still, flat vector logo.
• Lighting — the single highest-leverage phrase: "soft window light", "Rembrandt lighting", "golden-hour backlight".
• Camera / framing — lens and shot type: "85mm portrait, shallow depth of field", "wide establishing shot", "vertical 9:16".

Write your first prompt step by step

Start from a real subject and build outward. Keep it to one or two sentences — models follow a focused prompt better than a paragraph of stacked keywords.

1. Name the subject and action. Begin concrete: "A young woman laughing while holding a bouquet of sunflowers." Specificity here anchors everything else.
2. Add one style. Pick a single dominant look — "editorial fashion photograph" — rather than stacking three competing styles, which muddies the output.
3. Describe the light. Name a lighting setup: "soft overcast daylight" or "warm rim light from behind". Lighting does more for realism than any adjective.
4. Set the camera. Add lens and framing: "35mm, waist-up, slightly low angle". This controls composition and depth of field.
5. Generate, then change one thing. Run it, then adjust a single variable — outfit, colour, time of day — and regenerate. Iterating one change at a time is how you learn what each word does.

What makes AI images look generic (and how to fix it)

Generic results almost always come from vague, model-pleasing words instead of specific description. The word "beautiful" tells the model nothing; "freckled skin, natural texture, soft key light" tells it exactly what to render.

Avoid "flawless" and "perfect" for people — they push models toward a plastic, over-retouched look. Name the film stock or lighting you actually want instead. And when you need legible text in the image, quote the exact words in the prompt; GPT Image 2 and Nano Banana Pro render short quoted text reliably.

The fastest way to learn: start from a working prompt

You do not have to write prompts from scratch. The quickest way to improve is to copy a prompt that already produced a great image, then swap the subject while keeping the style, lighting and camera wording intact. That "keep the recipe, change the subject" habit is how most creators work.

Every image in the PicGens gallery shows the exact prompt that made it, so you can copy a proven recipe and adapt it in seconds.

Frequently asked questions

How long should an AI image prompt be?

One to three focused sentences is ideal. Models follow a specific, well-structured prompt better than a long list of keywords. Say the subject, style, lighting and camera clearly, then stop.

Do the same prompts work across different AI models?

The descriptive core (subject, style, lighting, framing) transfers well between GPT Image 2, Nano Banana and Midjourney. Model-specific parameters — like Midjourney's --ar or --sref — do not, so drop those when moving a prompt.

How do I get the same character or style across several images?

Keep the style and lighting block identical and only change the action or setting, or use a reference image in a model that supports it (GPT Image 2, Nano Banana). For Midjourney, reuse the same --sref style code.