youcutai

How to Write Prompts for an AI Image Generator

Updated Sep 3, 2026 · 9 min read

The anatomy of a prompt that works

Modern image models read prompts more like a brief than a keyword list. The tag-soup style of 2022 ("masterpiece, 8k, trending on artstation") no longer helps and can hurt. What works in 2026 is a clear description in roughly this order:

  1. Subject: who or what, with one or two defining details.
  2. Action or pose: what the subject is doing.
  3. Setting: where, and what is around it.
  4. Lighting: the single biggest lever on mood. Golden hour, overcast, studio softbox, neon, candlelight.
  5. Camera and composition: close-up, wide shot, 35 mm, shallow depth of field, eye level, top-down.
  6. Style and medium: photograph, oil painting, flat vector illustration, 3D render, watercolour.
  7. Colour and mood: muted earth tones, high-contrast, pastel, moody.

A complete example for the AI Image Generator:

A ceramic coffee mug with a matte terracotta glaze, sitting on a linen napkin on a rustic oak table, steam rising, soft window light from the left, shallow depth of field, 50 mm photograph, warm muted tones.

That is 40 words. Most good prompts are 25 to 60 words. Shorter prompts leave too much to chance; longer ones bury the important parts.

Match the prompt to the model

The generator offers five models, and they respond to prompts differently.

  • FLUX Schnell (1 credit) is fast and forgiving. Use it to iterate on composition. It ignores very long prompts, so keep it under 40 words.
  • FLUX Pro 1.1 (3 credits) is the photorealism workhorse. It rewards camera language (lens, aperture, film stock) and precise lighting descriptions.
  • Seedream 4 (3 credits) has strong aesthetics and 2K output. It handles stylised and editorial looks well and likes colour and mood words.
  • Nano Banana (3 credits) excels at consistent characters across multiple images and at clean text rendering. Describe the character once in detail and reuse that exact paragraph in every prompt.
  • GPT Image (5 credits) has the best instruction following and typography. Write to it in full sentences, including layout instructions like headline at the top in bold sans-serif, product centred, logo bottom right. It is the right choice for posters, ads and anything with words.

A cheap workflow: draft with Schnell until the composition is right, then paste the same prompt into Pro, Seedream or GPT Image for the final. The free preview lets you compare outputs side by side before downloading.

Lighting, camera and style vocabulary

These words move the result reliably. Pick one or two from each group; combining five lighting terms produces mud.

Lighting: golden hour, blue hour, overcast soft light, hard midday sun, window light, rim light, backlit with lens flare, studio softbox, three-point lighting, neon signage, candlelight, volumetric fog.

Camera: extreme close-up, macro, medium shot, full body, wide establishing shot, top-down flat lay, low angle, eye level, 24 mm, 50 mm, 85 mm portrait lens, f/1.8 shallow depth of field, f/11 everything in focus, long exposure, tilt-shift miniature.

Medium and style: documentary photograph, editorial fashion photo, product photography on seamless white, cinematic still, Polaroid, 35 mm film grain, oil painting, gouache, watercolour, ink sketch, flat vector, isometric, low-poly 3D, claymation, pixel art, ukiyo-e, art deco poster.

Colour and mood: monochrome, duotone teal and orange, pastel, earthy, desaturated, high key, low key, moody, vibrant, muted.

Material: brushed aluminium, frosted glass, worn leather, raw linen, glossy ceramic, matte plastic. Naming materials is the fastest way to make product renders look real.

Negative prompts, aspect ratios and seeds

Negative prompts tell the model what to avoid. FLUX and Seedream support a separate negative field; GPT Image and Nano Banana prefer the exclusion written into the main prompt (no text, no people). Useful negatives: text, watermark, logo, extra fingers, blurry, low contrast, oversaturated, cropped head, duplicate subject. Keep the list short; a 30-item negative prompt fights the positive one.

Aspect ratio should be set before generating, not by cropping afterwards. 1:1 for feed posts and icons, 4:5 for Instagram portrait, 9:16 for stories and phone wallpapers, 16:9 for web heroes and slides, 3:2 for print. The model composes differently for each ratio, so a 16:9 landscape prompt will place the subject off-centre with negative space, which is what you want.

Seeds make results reproducible. When a generation is almost right, lock the seed and change one word at a time. This is how you learn what each word does. When you want variety, let the seed randomise and generate four at once.

Resolution: generate at the model's native size (1024 to 2048 px) and use the Image Upscaler for print. Asking a model for 8K in the prompt does nothing.

Copy-paste templates

Replace the bracketed parts.

Product on white (e-commerce) [Product] with [material and colour], centred on a seamless pure white background, soft studio lighting with a gentle shadow beneath, front three-quarter view, 50 mm, sharp focus throughout, no text, no props.

Lifestyle product shot [Product] placed on [surface] in a [room or outdoor setting], [time of day] light from [direction], shallow depth of field, editorial photograph, [colour palette], no people, no text.

Portrait / headshot Professional headshot of a [age] [gender] with [hair, features], wearing [clothing], neutral grey studio background, soft key light with subtle rim light, 85 mm lens, f/2, relaxed confident expression, natural skin texture.

Poster with text (GPT Image) A [event type] poster, [aspect ratio]. Headline reads "[TEXT]" in bold condensed sans-serif at the top. Below it, [illustration description]. Colour palette [colours]. Footer with date "[DATE]" and venue "[VENUE]" in small caps. Clean, [style] design.

Consistent character (Nano Banana) Character sheet: [name], a [age] [description: hair, eyes, build, outfit, distinctive accessory]. Same character in every image. Scene: [action and setting]. Flat illustration, clean lines, [palette].

Illustration / icon Minimal flat vector icon of [object], single [colour] on a white background, rounded geometry, no gradients, no text, centred, generous padding.

Save the ones you use as prompt presets inside the generator so a whole campaign shares one style.

Editing instead of regenerating

When an image is 90 % right, do not roll the dice again. Open it in the AI Photo Editor and describe the change: make the mug blue, replace the background with a beach at sunset, remove the second chair, add a small plant on the left. The editor keeps everything you did not mention and changes only what you asked for. Nano Banana (3 credits) is fast and preserves identity; GPT Image (5 credits) handles complex multi-step instructions and text edits.

Editing prompts follow their own rules: be specific about what and where, name what must stay the same (keep the face and lighting identical), and make one change per instruction. Chain several small edits rather than one long request.

This generate-then-edit loop is how professionals work now: broad strokes from the generator, precise control from the editor, and the upscaler at the very end for delivery.

Frequently asked questions

Should I still use words like "8k" and "masterpiece"?+

No. Current models ignore or are confused by quality tags. Describe lighting, camera and material instead; that is what actually raises perceived quality.

Which model is best for text in images?+

GPT Image for layouts and longer text, Nano Banana for short labels and consistent characters. Always put the exact text in quotation marks in the prompt.

How do I get the same character in several images?+

Write a detailed character description once, reuse it verbatim in every prompt, use Nano Banana, and lock the seed where possible. Alternatively generate one image and use the AI Photo Editor to change only the scene.

How much does a generation cost?+

FLUX Schnell is 1 credit; FLUX Pro, Seedream 4 and Nano Banana are 3; GPT Image is 5. On a Pro plan ($24.99 for 800 credits) that ranges from about $0.03 to $0.16 per image.

← All guides

Tools mentioned