A vague prompt gets you a vague image. The gap between a beginner and an expert in image generation is not the tool — it is the prompt. Every leading image model, from Midjourney to Google's Nano Banana Pro, responds to the same underlying grammar: specific subjects, layered details, and explicit constraints. Learn that grammar once and it works everywhere.
This guide breaks down a proven five-layer prompt structure, shows how to adapt it to each major model in 2026, and gives you copy-paste templates you can start using today.
The five-layer prompt structure
Nearly every model's official documentation converges on the same skeleton. Google's prompting tips for Nano Banana Pro and OpenAI's guidance for GPT Image models both recommend ordering a prompt as subject, action, setting, composition, then style and constraints. Ideogram even formalizes this into JSON fields.
Treat every prompt as five layers, written in this order:
1. Subject — who or what is in the frame, described with concrete nouns, not vague categories.
2. Action or pose — what the subject is doing, including body position and expression.
3. Setting and context — where the scene takes place, time of day, surrounding environment.
4. Composition and technical spec — camera angle, lens feel, lighting, aspect ratio.
5. Style and constraints — the medium (photo, 3D render, watercolor), plus what must not appear.
Here is the difference it makes in practice:
- Weak: "a businesswoman in an office"
- Strong: "a woman in her 40s wearing a charcoal blazer, reviewing a printed report at a standing desk, in a sunlit corner office with floor-to-ceiling windows, shot from a low three-quarter angle with soft window light, photorealistic, shallow depth of field, no text or logos visible"
The weak prompt leaves all five layers to chance. The strong one gives the model almost nothing to guess about. Organized, layered detail consistently outperforms both terse prompts and rambling, unstructured ones of the same length.
Layer 1: Be specific about the subject
The subject is the foundation. If it is vague, everything else falls apart.
- Weak: a car
- Better: a red sports car
- Strong: a red Ferrari 250 GTO driving at high speed, gleaming chrome details
Describe age, clothing, materials, and distinguishing features. But keep the subject description focused — ten precise words beat thirty meandering ones. Do not overload this layer; the other four layers exist to carry the remaining detail.
Layer 2: Describe the action or pose
AI image models need verbs, not just nouns. Action pins down body position, expression, and energy.
Instead of "a runner", write "a runner mid-stride, leaning into the final lap, sweat on their brow, determined expression". Instead of "a dog", write "a golden retriever puppy sitting attentively, head tilted, tongue out".
Action is also where you handle multi-subject scenes: "two chefs tossing a pizza dough between them while a crowd watches". Name who does what.
Layer 3: Build the setting
The environment is where most beginners fall short. The model does not just need *what* — it needs *where*.
- Weak: a red sports car
- Better: a red sports car on a city street
- Strong: a red sports car driving through a rainy neon-lit Tokyo street at night, reflections on wet asphalt
Setting carries time of day, weather, mood, and spatial context. Add two to three concrete environmental details: light sources, background elements, atmosphere. The setting is the single biggest lever for making an image feel cinematic rather than flat.
Layer 4: Composition, lighting, and camera terms
Lighting is the most skipped prompt element, and the most transformative. Terms like "golden hour", "soft studio lighting", "dramatic rim lighting", and "neon ambient glow" instantly shift the mood.
Useful camera and composition keywords:
- Shot type: close-up, wide shot, aerial view, macro
- Angle: low angle, overhead, three-quarter view, eye-level
- Depth: shallow depth of field, deep focus, bokeh background
- Aspect: 16:9 landscape, vertical 9:16, square 1:1
Stick to one or two composition cues per prompt. Conflicting camera instructions — "close-up wide shot" — confuse the model.
Layer 5: Style plus what to exclude
Name one style, at most two. "Watercolor illustration" or "cinematic photorealism" gives the model a clear target. Five styles stacked together usually hurt results because the model cannot resolve them.
The second half of this layer is the negative prompt: explicitly state what must not appear. "No text, no watermarks, no logos" is the classic trio. For photorealism add "no deformed hands, no extra fingers". Most modern models (Midjourney V7, Nano Banana Pro, Seedream 4.5) accept negative instructions inside the prompt itself rather than a separate field.
Model-specific adjustments for 2026
The five layers work everywhere, but each leading model has quirks worth knowing.
Nano Banana Pro (Google): its strength is photorealism and character consistency across edits. Emphasize logical consistency with words like "photorealistic", "consistent lighting", and "accurate proportions". It handles reference images and iterative editing well, so you can generate a base scene and refine it with follow-up instructions instead of rewriting the whole prompt.
Seedream 4.5 (ByteDance): its standout skill is text rendering and multi-image layouts. Give it explicit typography instructions: "Title 'Summer Collection' in elegant serif font at top left, product centered, price '$29.99' below". This is the model to pick when your image must contain readable text.
GPT Image 1.5 (OpenAI): the most artistically flexible of the bunch. Explore art movements and media freely: "art nouveau poster design", "in the style of a Hayao Miyazaki film", "gouache painting". It also supports reference images for style transfer.
Midjourney V7/V8: optimized for aesthetic quality and concept art. Keep prompts a little shorter than you would for Nano Banana — Midjourney responds well to comma-separated style fragments, and it supports reference parameters (sref, cref) for locking a style or character across generations.
Ideogram: formalizes prompting into JSON fields for description, style, and compositional breakdown. If you are building image generation into a product, structured JSON prompts on Ideogram are easier to template programmatically.
Five copy-paste prompt templates
Product photography:
"matte black wireless headphones on a concrete pedestal, studio product shot, dark charcoal background with subtle gradient, soft overhead studio lighting, dramatic shadows, shallow depth of field, photorealistic, premium e-commerce style, no text, no logos visible"
Portrait:
"a woman in her 30s with freckles and curly auburn hair, smiling softly, golden hour light in a sunflower field, gentle breeze moving her hair, shot at eye level with a shallow depth of field, cinematic photorealism, warm tones, no text"
Blog hero image:
"flat vector illustration of a marketer reviewing ranking charts on a laptop, modern home office with plants and a large window, teal and orange color palette, soft ambient lighting, minimalist geometric style, wide 16:9 composition, no text, no watermarks"
Fantasy concept art:
"an ancient stone lighthouse on a cliff during a thunderstorm, waves crashing against the rocks, a hooded figure holding a lantern at the top, dark fantasy style, dramatic lightning illumination, moody atmosphere, cinematic wide shot, no text"
Marketing poster with text (Seedream 4.5):
"summer beach party poster, retro sunset gradient, headline 'SUMMER SALE' in bold condensed sans-serif at the top, 50 percent off badge in a circle at bottom right, palm trees and ocean waves, vibrant colors, no distorted text, no watermark"
Iterate like a pro: change one thing at a time
When a result misses, do not rewrite the entire prompt. Identify which layer failed — was the subject wrong, the lighting off, or the style off-target? — and change only that layer.
Keep a simple prompt log: the prompt you used, the model, and a one-line note on what to adjust. After a dozen generations you will have a personal library of what works. Many prompt engineers also save successful prompts as reusable templates with a placeholder for the subject, so a proven portrait formula can be reused for a completely different person.
One more habit that separates professionals: generate in small batches of two to four variations per prompt rather than one. Slight variations reveal which details the model interprets well and which it ignores, and you can then lock the good ones into the prompt explicitly.
Key takeaways
- Structure every prompt in five layers: subject, action, setting, composition, style plus constraints.
- Be concrete — name materials, colors, camera angles, and times of day instead of abstractions.
- Name one style (two at most) and always include negative prompts like "no text, no watermarks".
- Adapt to the model: Nano Banana Pro for photorealism and consistency, Seedream 4.5 for text in images, GPT Image 1.5 for artistic range, Midjourney for aesthetics.
- Iterate by changing one layer at a time and keep a prompt log of what works.