How to write prompts for AI images

Prompt structure

Lead with what matters most. FLUX gives more weight to words at the start, so put the subject and action first, then style and context, and minor details last. Most tasks need 30 to 80 words, complex scenes more.

Purpose helps too: "a cover image for an article about coffee" tells the model more about composition and mood than just "coffee".

• Purpose: what the image is for
• Subject: who or what is in the frame, specifically
• Action and pose
• Place and time
• Style: photo, illustration, 3D, painting
• Lighting and color
• Composition: shot size, angle, empty space

Style and lighting

For photorealism, use a photographer’s vocabulary: lens, focal length, aperture, film stock. Models treat these as cues for the look rather than an exact camera simulation. For illustration, name the technique and the visual traits: outlines, shading, palette.

Convey mood with details rather than words like "atmospheric". "Wet asphalt reflects a neon sign, light fog" is more precise than "cyberpunk vibe".

Composition and format

Say where the subject sits and what surrounds it: "product centered, empty space at the top for a headline". This matters most for banners and covers that will carry text later.

Set the aspect ratio with the model parameter rather than words in the prompt. A phrase like "vertical photo" without the matching setting is often ignored.

Text in images

Put the exact words in quotes and describe placement, font, size and color. Spell unusual words and brand names letter by letter. Ask for no other text, and always check the spelling in the result.

GPT Image and Nano Banana Pro handle text best. Small type and Cyrillic still produce errors, so important lettering is safer to add in the design editor on top of the image.

• "Coffee shop poster, the word “MORNING” in large white sans serif in the center, a cup on a wooden table below, warm light"
• "A red neon sign reading “OPEN” above a door, night street, reflections on wet tiles"

How Flux, Seedream, Nano Banana and GPT Image differ

The general rules are the same, but each model has its own strengths and a preferred prompt style.

Negative prompts and iteration

FLUX does not support negative prompts, and Google recommends describing what you want instead of what to avoid. Write "sharp focus throughout" instead of "no blur" and "an empty street" instead of "no people".

Improve in steps: one change at a time, using the best variant as the base for the next. If the model starts drifting, restate the key constraints instead of saying "same as before".

How to do it in Solistra

The Image section brings together FLUX, Seedream, Nano Banana, GPT Image, Ideogram, Recraft and other models. The credit price updates as you change the model, quality and number of variants.

• Open Image and choose a model for the task
• Write the prompt using the structure and set the format as a parameter
• Generate several variants and pick the best
• Refine it with a text edit or in the design editor