How to Write AI Image Generation Prompts

Fundamentals of AI Image Prompting

AI image generation prompts are natural language instructions passed to generative models—such as Midjourney, Stable Diffusion, Black Forest Labs' Flux, or OpenAI's DALL-E 3—to direct the synthesis of visual scenes. At their core, these models do not "understand" art, narrative, or photography in human terms. Instead, they pair deep text encoders (such as CLIP or T5) with diffusion architectures that learn cross-modal associations between descriptive text tokens and visual patterns across hundreds of millions of training images. Prompt Basics How to Write AI Image Prompts - Beginner to Advanced Guide

Writing an effective AI prompt requires translating an abstract creative idea into concrete, semantically distinct descriptions. While human communication often relies on implicit assumptions, diffusion models construct imagery token by token based on statistical probabilities. If a prompt omits details regarding lighting, medium, camera perspective, or background context, the model fills those semantic voids with its internal training biases. Precision, word order, and semantic clarity govern the degree of control an artist maintains over the output. How to Write AI Image Prompts - Beginner to Advanced Guide How to Write AI Image Prompts: The 6-Part Formula (2026)

The Anatomy of an Effective Image Prompt

High-performing prompts generally follow a structured hierarchy. While natural language syntax varies by model, decomposing a concept into modular building blocks ensures that essential visual variables are explicitly defined rather than left to chance. How to Write AI Image Prompts - Beginner to Advanced Guide How to Write AI Image Prompts: The 6-Part Formula (2026)

[Core Subject] + [Environment/Setting] + [Medium & Artistic Style] + [Lighting & Color Palette] + [Composition & Framing] + [Technical Parameters]

1. Subject Definition and Action

The subject is the focal point of the composition. Specificity prevents generic generations:

• Vague: a warrior
• Specific: an elderly Nordic blacksmith with a braided gray beard, wearing a charred leather apron and iron armguards, hammering a glowing red ingot

Specify materials, physical condition, pose, expression, and active verbs. Action anchors the subject dynamically within the frame rather than producing a static, mannequin-like pose. How to Write AI Image Prompts: The 6-Part Formula (2026)

2. Environment and Spatial Context

Backgrounds require active definition to prevent diffusion models from defaulting to blurred, indistinct studio backdrops:

• Spatial relationship: Define foreground, midground, and background elements (e.g., “in the foreground, wet cobblestones reflect streetlamps; in the background, towering Gothic spires vanish into thick mist”).
• Atmospheric conditions: Weather, haze, airborne particulates (dust motes, embers, falling snow), and time of day establish physical depth. How to Write AI Image Prompts - Beginner to Advanced Guide How to Write Image Prompts (Midjourney, DALL·E, SD)

3. Medium, Materiality, and Artistic Style

Diffusion models synthesize photographic textures differently from illustrative or painterly mediums. Clearly specifying the medium grounds the model's rendering engine:

• Photographic: “Editorial portrait, 35mm film photograph, candid street photography, macro detail.”
• Fine Art & Traditional Media: “Impasto oil painting with visible palette knife strokes,” “woodblock print with rough linework,” or “wet-on-dry watercolor illustration with subtle pigment pooling.”
• Digital & Graphic: “Vector illustration, flat geometric shading, bold outlines,” or “isometric 3D digital render, matte clay finish.” How to Write AI Image Prompts - Beginner to Advanced Guide How to Write Image Prompts (Midjourney, DALL·E, SD)

4. Lighting and Color Science

Lighting dictates mood, highlights structural geometry, and determines tonal contrast. Avoid generic adjectives like "beautiful lighting" in favor of technical descriptors:

• Direction and Source: Volumetric god rays, directional rim lighting, chiaroscuro contrast, soft diffuse daylight, bioluminescent rim glow, neon sign spill.
• Color Temperature and Palette: Monochromatic with emerald accents, warm golden hour amber tones, desaturated muted pastels, split-complementary teal and orange. How to Write AI Image Prompts - Beginner to Advanced Guide How to Write AI Image Prompts: The 6-Part Formula (2026)

5. Camera Optics and Composition

Instructing the generator using photographic terminology aligns the latent vectors with real-world camera compositions:

• Shot Type: Extreme close-up, low-angle hero shot, Dutch tilt, wide panoramic landscape, bird's-eye view.
• Optics & Depth: Shot on 85mm portrait lens, f/1.4 shallow depth of field, sharp foreground focus with creamy bokeh background, 24mm wide-angle perspective distortion. How to Write AI Image Prompts: The 6-Part Formula (2026) How to Write Image Prompts (Midjourney, DALL·E, SD)

Model Architectures and Prompting Paradigms

Different AI generators employ distinct text-processing frontends. A prompt optimized for one engine will often fail or produce unintended results when used directly in another. How to Write AI Image Prompts - Beginner to Advanced Guide How to Write Image Prompts (Midjourney, DALL·E, SD)

Conversational Prose vs. Comma-Separated Token Soup

Early generation models (such as Stable Diffusion 1.5) relied heavily on "tag soup"—chains of comma-separated keywords like masterpiece, trending on artstation, 8k, highly detailed, sharp focus. Modern architectures like Flux.1 and DALL-E 3 utilize massive multilingual language models (such as T5-XXL) that interpret grammar, spatial prepositions, and structural relations. How to Write AI Image Prompts - Beginner to Advanced Guide How to Write AI Image Prompts: The 6-Part Formula (2026)

• Legacy Token Pattern (SDXL / SD 1.5):

  cyberpunk alley, neon lighting, rainy, reflections, highly detailed, 8k, cinematic, unreal engine 5 render, wide shot

• Modern Natural Language Pattern (Flux / DALL-E 3 / Midjourney v6):

  A wide cinematic photograph of a narrow cyberpunk alleyway at midnight during a downpour. Vibrant magenta and cyan neon signs reflect across standing puddles on the wet asphalt. A lone courier wearing a yellow transparent raincoat walks away from the camera into the distance.

Modern models parse words like "reflect across," "during," and "wearing" to create coherent spatial relationships, whereas legacy tag collections frequently jumble these elements across the canvas. How to Write AI Image Prompts - Beginner to Advanced Guide How to Write AI Image Prompts: The 6-Part Formula (2026)

Advanced Prompt Engineering Techniques

1. Token Weighting and Emphasis

In models like Stable Diffusion, artists can adjust the mathematical influence of specific words:

• In Automatic1111/ComfyUI syntax, (word:1.2) increases weight by 20%, while [word] or (word:0.8) reduces it.
• In Midjourney, multi-prompt separators assign numeric weights: steampunk airship::2 stormy sky::1 sunset::0.5. This prevents high-concept elements from being drowned out by background tokens. Prompt Basics How to Write AI Image Prompts - Beginner to Advanced Guide

2. Negative Prompting

Negative prompts instruct the model on what features to penalize or omit in latent space. Rather than cluttering the positive prompt with negations (which models often misunderstand due to lack of negation parsing in early CLIP encoders), negative prompts isolate unwanted traits:

• Midjourney: Handled via the --no flag (e.g., --no blur, text, frame, oversaturated). Prompt Basics
• Stable Diffusion: Dedicated negative field (e.g., blurry, low-resolution, mutated hands, extra limbs, watermark, cartoonish, oversaturated).

3. Aspect Ratio and Model Parameters

Most engines support command-line arguments that adjust mechanical output settings without altering the semantic prompt:

• Aspect Ratios: Default squares ($1:1$) frequently clip environmental scenes. Midjourney uses --ar 16:9 for widescreen cinematic landscapes or --ar 4:5 for social portraiture. Flux and SDXL accept exact pixel dimensions (e.g., $1344 \times 768$ or $832 \times 1216$). Prompt Basics How to Write AI Image Prompts - Beginner to Advanced Guide
• Stylize / Variance: Midjourney offers parameters like --stylize [0-1000] (governing adherence to prompt versus native aesthetic training) and --weird [0-3000] (introducing atypical latent noise). Prompt Basics
• Seeds: Specifying a fixed random seed (--seed 42 or manual seed entry in WebUIs) stabilizes noise initialization, enabling controlled testing of incremental prompt revisions. Prompt Basics How to Write AI Image Prompts - Beginner to Advanced Guide

Common Prompting Pitfalls and Solutions

Negative Word Traps ("Pink Elephant Paradox")

Diffusion models struggle with syntactic negation in positive text strings. If a prompt says "a living room with no furniture", the text encoder extracts the token furniture, statistically elevating the probability of chairs and tables appearing.

• Correction: State the desired presence rather than the absence: "an empty, unfurnished living room with bare wooden floorboards and exposed plaster walls." How to Write AI Image Prompts - Beginner to Advanced Guide

Quality "Buzzword" Inflation

Adding strings like photorealistic, hyperrealistic, 4K, 8K, octane render, trending on ArtStation to modern models often degrades output quality. In modern datasets (especially Flux and Midjourney v6), professional photographic assets are rarely tagged with terms like "photorealistic" in curated metadata; instead, such tags frequently appear in amateur forum uploads or low-fidelity 3D renders. How to Write AI Image Prompts - Beginner to Advanced Guide How to Write AI Image Prompts: The 6-Part Formula (2026)

• Correction: Replace buzzwords with concrete technical specifications: mention sensor sizes ("medium format photography"), film stocks ("Fujifilm Pro 400H, fine grain"), or lighting types ("softbox diffusion, rim lighting"). How to Write AI Image Prompts - Beginner to Advanced Guide

Concept Bleeding

When a prompt requests multiple colored objects (e.g., "a man in a red jacket standing next to a blue car"), CLIP-based encoders often cross-contaminate tokens, producing a man in a blue jacket or a red car.

• Correction: Group descriptors tightly, use spatial prepositions, or leverage modular workflows such as regional prompting, Inpainting, or ControlNet layers to isolate subjects spatially. How to Write AI Image Prompts - Beginner to Advanced Guide

Practical Prompt Walkthrough

The following progressive iterations illustrate how refining composition, medium, and lighting transforms an image concept from a generic baseline into an intentional, production-ready output.

Concept: A botanical glasshouse at night

• Iteration 1 (Basic Subject):

  a greenhouse with plants at night

  • Result: Static framing, flat ambient darkness, generic houseplants, lack of atmosphere.

• Iteration 2 (Adding Environment and Architectural Medium):

  A Victorian cast-iron conservatory filled with exotic tropical palms and hanging ferns, exterior night view, glass panels showing interior foliage

  • Result: Clear architectural style and distinct botanical elements, but flat lighting.

• Iteration 3 (Adding Lighting, Atmosphere, and Palette):

  A sprawling Victorian cast-iron glasshouse surrounded by dense mist at twilight. Warm golden light emanates from within the structure, illuminating lush monsteras and towering banana palms. Cool indigo dusk sky with faint stars visible through the glass.

  • Result: High tonal contrast, complementary color temperature (warm gold versus indigo), and atmospheric depth.

• Iteration 4 (Final Production Prompt with Camera Optics and Style):

  Architectural photograph of a grand Victorian glass conservatory at dusk. Warm incandescent lighting glows from inside, casting sharp structural shadows across damp stone gravel paths outside. Lush tropical foliage, dripping ferns, and giant monsteras are visible through rain-streaked glass panes. Ground-level wide shot, 28mm architectural lens, balanced exposure, deep shadows, cinematic mood. --ar 16:9 --style raw

  • Result: Fully realized scene with controlled perspective, defined textures (rain-streaked glass, damp gravel), balanced exposure, and an intentional cinematic aspect ratio. Prompt Basics How to Write AI Image Prompts - Beginner to Advanced Guide How to Write Image Prompts (Midjourney, DALL·E, SD)

What a Good Image Prompt Actually Contains

Writing an effective AI image prompt means describing a visible scene rather than stating an abstract idea, and doing it in a predictable order: subject → action or pose → setting → composition and camera → lighting → medium and style → technical parameters. Google's own guidance for its image models expresses this as a formula of subject, action, location/context, composition, and style, with worked examples like a fashion model in a tailored brown dress, posing with a confident stance, against a seamless deep cherry red studio backdrop, framed as a centered medium-full shot. Midjourney's documentation, by contrast, advises that short and simple prompts generally produce its best images and recommends clear phrases that describe what you want to see. Both are correct — for their own systems — which is the first thing to understand about prompting: there is no single universal prompt grammar, only conventions that each model rewards.

A practical way to think about it: the model has no access to your intent, only to the words. Every attribute you leave unspecified gets filled in from statistical defaults in the training data — usually a centered subject, neutral lighting, generic "digital art" rendering, and whatever demographic or stylistic average the model has absorbed. Prompting is therefore less about finding magic words and more about deciding which defaults you want to override.

The Slots Worth Filling

Rather than memorizing a template, it helps to treat a prompt as a set of optional slots. Fill the ones that matter for your image and leave the rest for the model to decide.

OpenAI's image-prompting guidance emphasizes naming materials, lighting, colors, and the visual medium; requesting "photorealistic" or "real photograph" explicitly when that is the goal; and specifying body framing, gaze, and interaction with objects through instructions such as "full body visible, feet included" or "hands naturally gripping the handlebars." It also notes something easy to miss: camera specifications are cues for appearance, not guarantees of physical simulation — writing 85mm f/1.4 nudges the model toward images that look like they were shot that way, it does not run an optical model. For wide, cinematic, low-light, or neon scenes, the advice is to specify scale, atmosphere, and color rather than relying on mood words alone.

Two further pieces of guidance from Google's best-practices documentation are worth internalizing because they contradict common habits. First, describe what you want rather than what you don't — instead of "no cars," write "an empty, deserted street with no signs of traffic." Second, use photographic and cinematic vocabulary to control the camera, with terms like "wide-angle shot," "macro shot," or "low-angle perspective." For complex scenes, the same guidance suggests splitting the request into sequential steps: establish the background first, then add the foreground element, then the focal object.

Here is what filling the slots looks like in practice, written as a single instruction-style prompt of the kind that suits modern text-following models:

A photorealistic candid photograph of an elderly sailor on a small fishing
boat, calmly adjusting a net while his dog sits nearby on the deck.
Weathered skin with visible wrinkles and pores. Shot like a 35mm film
photograph, medium close-up at eye level, 50mm lens. Soft coastal daylight,
shallow depth of field, subtle film grain, natural color balance. Honest and
unposed, with real skin texture and worn materials. No heavy retouching.

Note that this prompt specifies not only content but failure modes to avoid stylistically — "no glamorization, no heavy retouching" — which is how you counteract the glossy, over-rendered default that many models drift toward. OpenAI's cookbook examples use exactly this pattern, and apply related logic to other categories: for logos, favoring simplicity over detail so the mark reads at small and large sizes, with flat shapes, minimal strokes, and generous padding.

Keyword Dialects Versus Instruction Dialects

Models differ in how they parse language, and this is the single most common source of frustration when a prompt that worked in one tool produces nonsense in another.

Keyword-weighted systems — the classic Stable Diffusion family and, to a large extent, Midjourney — respond well to compact noun phrases and comma-separated descriptors, and tend to weight earlier tokens more heavily. Midjourney guidance commonly describes moving from what the image is to how it should look: subject first, then context and composition, then style or medium, so that the words carry the structure and parameters only fine-tune. Practitioner guides for Midjourney frequently note that it prefers descriptive noun phrases over conversational sentences.

Instruction-following systems — image models fronted by a large language model, including recent OpenAI and Google offerings — handle full sentences, conditional logic, exact quoted text, and multi-step requests far better. With these, writing a paragraph of plain English often outperforms a keyword soup. Google's documentation even notes the practical necessity of phrasing like "create an image of" or "generate an image of," since a multimodal model might otherwise reply with text instead of an image.

For rendering words inside an image, instruction-style models are generally the better tool, and the technique matters: put the required wording in quotation marks, describe its position and typography, spell unusual names letter by letter if needed, ask for no extra text, and then verify spelling and legibility in the output. For charts and diagrams, include the actual numbers and labels in the prompt and check factual relationships, not just appearance.

Parameters, Seeds, and Reproducibility

Beyond the prose, most tools expose numeric controls. In Midjourney these are flags appended with a double dash at the end of the prompt — aspect ratio (--ar), model version (--v), quality (--q), randomness (--chaos), and exclusions (--no), among others. Two dials are often confused: stylize governs how strongly the system applies its own house aesthetic versus a literal reading of your words, while chaos governs how different the images in one grid are from each other, affecting spread rather than the style of any single image. Low chaos suits refinement; high chaos suits early exploration. Exact ranges, defaults, and available flags change between versions, so treat the provider's current documentation as authoritative rather than any cheat sheet.

The seed is the most underused control for serious work. In diffusion systems, the seed is the integer that determines the starting noise pattern; the same seed with the same prompt on the same model version reproduces the same image. That makes controlled experiments possible: fix the seed, change exactly one phrase, and you can attribute the difference to that phrase rather than to randomness. A documented workflow for one user involved holding a seed constant while swapping a single name in the prompt to isolate its effect.

Negative Prompts and Guidance Strength

A negative prompt is a second text field listing what the image should not contain. It is not a post-hoc filter. Under classifier-free guidance, the model makes two predictions at each denoising step — one conditioned on your prompt, one on an unconditional (normally empty) input — and the sampler steers away from the second toward the first:

where $s$ is the CFG (guidance) scale. A negative prompt replaces that unconditional term, which is why it shapes the image from the very first step and why its strength is proportional to $s$. With guidance at 1, a negative prompt does essentially nothing by construction.

Guidance scale is a trade-off rather than a quality slider. Low values give loose, more creative interpretations; very high values push the difference so hard that output turns oversaturated and artifact-heavy. Commonly cited working ranges sit in the mid single digits for many modern checkpoints, with lower values recommended for the newest architectures, but the appropriate range is model-specific and shifts with each release. If the model is ignoring your prompt, raising guidance sometimes helps — but more often the real problem is ambiguous wording, conflicting descriptions, or a concept the model never learned well. Conversely, when images show banding, crunchiness, or distortion, lowering guidance is an underrated first move.

Negative prompting conventions vary sharply by architecture, and copying a long boilerplate list between models is a reliable way to make things worse:

• Older checkpoints tolerate and sometimes benefit from longer negative lists addressing anatomy and quality.
• Larger, better-aligned models generally do better with short, targeted negatives aimed at specific unwanted objects, styles, or lighting rather than generic quality keywords.
• Distilled fast variants that run at very low guidance barely react to negatives at all.
• Some newer architectures were not designed around negative prompting, and certain systems use it only weakly — in which case sharpening the positive prompt is the better investment.

Where emphasis syntax is supported, weight only the one or two terms addressing your actual recurring defect; weighting everything is equivalent to weighting nothing, and pushing weights far above moderate levels tends to wash out color or flatten the image because the model is being driven away from features that overlap with ordinary photographs. Emphasis notation is not universal, so confirm what your interface supports.

Iterating Rather Than Rewriting

The habit that separates consistent results from lucky ones is changing one variable at a time. Start with a plain description and no parameters, generate, and see where the default lands; only then add controls to push in a specific direction. Reach for a parameter when words have stopped working, not before.

With conversational and reference-image-capable models, iteration often means editing rather than re-prompting: supplying the previous image and asking for a small, named change — warmer lighting, a more serious expression, a different background — while leaving everything else alone. Reference images also do work that description struggles with: character identity, a specific product, a brand palette, or a hard-to-articulate visual style. Provider documentation for image editing and reference-based generation includes templates for exactly these workflows, and instructions like "the background is white" or "no text" remain useful even in that mode.

A disciplined loop looks like this:

1. Write a one-sentence description of the subject and setting only. Generate four.
2. Pick the closest result and note specifically what is wrong — framing, lighting, medium, mood.
3. Fix the seed if available. Add or change one slot addressing the single biggest problem.
4. Add a negative prompt only for defects you have actually observed, keeping it short.
5. Once the image is close, adjust aspect ratio, guidance, and quality settings last.
6. Save the full prompt, model version, settings, and seed together — a prompt without its settings is not reproducible.

Limits Worth Planning Around

Some weaknesses are structural rather than prompt-fixable. Precise counting ("exactly seven birds"), legible small text, complex hand positions, consistent faces across separate generations, accurate anatomy in unusual poses, and technically correct diagrams all remain unreliable to varying degrees depending on the model, and asking more emphatically does not help. For diagrams and infographics, verify the labels and relationships, not just the look.

There are also non-technical constraints. Prompting in the name of a living artist to replicate their style raises ethical and, depending on jurisdiction, legal questions that are genuinely unsettled; the same applies to generating recognizable real people, trademarks, and copyrighted characters. Provider usage policies restrict some of this independently of the law, and those policies change. For commercial work, output rights, model-specific licensing terms, and disclosure obligations vary by provider, plan, and region — worth checking current terms directly and, for anything consequential, with qualified counsel. Describing the qualities you admire — brushwork, palette, line weight, period, printing process — is usually both more defensible and more controllable than invoking a name.