How to Write AI Prompts for Images: Guide, Tips and Examples

What an AI Image Prompt Tells the Model

An AI image prompt supplies specific mathematical coordinates inside a model's joint vision-language embedding space, defining subject, style, context and visual details. Dual text encoders, such as CLIP and T5, parse the input string to extract semantic relationships and visual attributes.

CLIP text encoders map concrete visual tokens (say "leather texture" or "studio spotlight") directly to visual feature clusters learned during pre-training, as documented in the large-scale prompt gallery research behind DiffusionDB (Wang et al., 2023). Each word of the prompt is tokenised and embedded into a fixed-dimensional vector that, together with the sampled noise tensor, steers the diffusion trajectory toward the requested image. Inside cross-attention, the text embeddings act as keys and values while the image latents act as queries.

Large language encoders like T5 handle long-range contextual relationships and complex syntax, which helps the underlying diffusion transformer interpret multi-object spatial positioning. Every adjective, noun and technical parameter modifies the cross-attention matrix, dictating where denoising steps allocate pixel density and visual weight in latent space. In plain terms: CLIP handles the visible, nameable things (materials, objects, colours); T5 handles relationships between them ("behind", "reflected in", "while holding").

Why Specific Language Produces Better Results

Descriptive, concrete language raises generation fidelity because explicit visual terms cut semantic noise and narrow the model's output distribution. Evaluative words such as "beautiful", "stunning", "hyperrealistic", "4K" or "5K" lack standardised mathematical representations in latent space and often degrade prompt adherence. They are judgments, not descriptions, so the encoder can only map them to an averaged blur of training data.

Empirical benchmarks confirm that replacing subjective hype with technical specification yields measurable gains in image-text alignment. Automatic prompt-engineering research on realistic image synthesis reports that appending optimal camera descriptions to a base prompt raises semantic consistency and text-image alignment relative to generic phrasing.

Prompt-engineering surveys reach the same conclusion from the taxonomy side: explicit task framing consistently outperforms underspecified phrasing across modalities (The Prompt Report: A Systematic Survey of Prompt Engineering Techniques, 2024). Practical translation of that finding: swap "a stunning portrait of a banker" for "a three-quarter studio portrait of a banker, 85mm lens, softbox key light, neutral grey backdrop". One of those strings is a wish. The other is a brief.

Start With the Main Subject and Action

The opening phrase of a writing prompt should state the central subject and its immediate physical action. Defining the core subject first secures primary attention tokens in the conditioning array before downstream modifiers are processed.

A prompt that starts with "A senior risk analyst reviewing financial ledgers" establishes clear subject semantics straight away. Dropping vague introductory filler like "Generate an image of" stops you wasting token slots on non-descriptive syntax. An explicit verb also suppresses the default "posing for a stock photo" stance that models fall back on when no action is supplied.

User-journey research on text-to-image tools documents the failure mode directly: participants who buried the subject at the end of long prompts hit unexpected outputs far more often, because the attention structure had already committed to earlier modifiers.

Governance documentation supports the same discipline from a risk angle: model use, assumptions and limitations should be documented so that ambiguity about what the model must represent is reduced before generation, not after publication (NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative AI Profile, 2024).

Add Style, Mood and Context

Once the subject is set, specify the surrounding environment, atmospheric mood and overall artistic direction. Contextual modifiers tell the model how to integrate the subject into its background instead of pasting it on top.

Environmental detail such as "inside a glass-walled conference room at twilight with soft ambient rain" anchors the subject in physical space. Medium tags like "architectural photography" or "minimalist vector art" enforce a consistent visual language across the whole generated image.

Thematic analysis of more than three million real-world prompts confirms that explicit style and mood tags improve output coherence compared with isolated subject nouns.

A short modifier taxonomy keeps this step fast:

Style / mediumwatercolour, charcoal sketch, oil painting, concept art, vector illustration, editorial photo, 3D render.

Mood / emotionserene, tense, melancholy, dreamlike, energetic, austere, documentary-neutral.

Scene contexturban street at dusk, glass-walled trading floor, mountain meadow after rain, under an oak tree.

Technical atmospheresoft diffused light, volumetric haze, hard specular highlights, shallow depth of field.

Put the Most Important Details First

Position critical requirements near the start of the string, because transformer positional encodings grant higher initial weight to early tokens. Token influence decays across extended sequences, so early positioning matters for non-negotiable scene attributes. Transformers have no innate notion of order; meaning depends on positional embeddings added to each token, which is why identical words placed differently produce different layouts (Attention Is All You Need, Vaswani et al., 2017).

Attention-map visualisation in interactive prompting research shows that early modifiers exert stronger influence over the central regions of the canvas than identical instructions appended at the tail.

Placing critical composition constraints, such as "Wide-angle aerial shot", at the front of the sequence makes the model establish structural framing before it renders localised surface textures. When a style must dominate everything else (single-line vector artwork, for instance), front-load two style phrases and repeat a short reinforcement tag at the tail: single line vector graphic, continuous black line, [subject], black and white.

Specify Composition, Framing and Perspective

Define camera viewpoint, shot distance and structural balance explicitly to control how subjects sit within the frame. Precise camera terminology replaces random visual arrangement with intentional composition.

Recognised optical terms such as "low-angle shot", "eye-level three-quarter profile" or "macro close-up with 85mm lens" establish exact spatial relationships. In an enterprise asset workflow, specifying "centered composition on neutral background" ensures generated graphics drop cleanly into pre-defined design templates.

A usable framing lexicon covers three axes:

Automatic prompt-engineering research quantifies why these tags matter: appending camera descriptions (angle, focal length, shot type) measurably improves both semantic consistency and text-image alignment versus generic wording.

Shot scaleextreme close-up, close-up, headshot, medium shot, full-body shot, wide shot, establishing shot.

Angleeye-level, low angle, high angle, overhead or top-down, underside view, Dutch tilt.

Orientation of the subjectfront view, side profile, three-quarter view, back view, over-the-shoulder.

Control Lighting, Color and Level of Realism

Direct lighting types, colour palettes and rendering fidelity to set the emotional tone and physical surface qualities of the image. Lighting parameters determine shadow hardness, highlight placement and material texture depth.

Specifying a studio setup such as "Rembrandt lighting with a high angled key light and warm fill" produces clear facial highlights and soft directional shadows: the key light sits high and off-axis, producing an illuminated triangle on the cheek opposite the light. Corpus analysis suggests these lighting tags are not decorative. They form stable thematic clusters across millions of prompts and correlate with higher-quality outputs.

Useful lighting and colour controls:

Colour parameters should define explicit palettes, such as "monochrome slate grey with warm amber accent lighting", rather than vague colour talk. For realistic outputs, "raw photo texture, subtle skin pores, natural daylight" yields better physical accuracy than subjective phrases like "hyperdetailed". Yes, that is less exciting to type. It works better.

Setupkey light, fill light, rim or hair light, background light, softbox, bounce, practical lights.

Qualityhard light (sharp shadows, high contrast) versus soft light (gradual transitions, low contrast).

Temperature2700 to 3000 K warm tungsten, 5500 K neutral daylight, 6500 K cool overcast.

Atmospherevolumetric light shafts, backlit haze, silhouette against window, golden hour, blue hour.

Choose Art Styles That Match the Intended Result

Select explicit artistic mediums and historical visual styles to guide rendering technique and aesthetic. Aligning style keywords with target outputs keeps execution consistent across campaign assets, which is the whole point when twelve images must look like one set.

Different visual mediums need distinct modifier triggers to activate the matching model capability:

• Photorealism: "35mm photograph, shot on Canon EOS R5, f/2.8 aperture, natural daylight, subtle film grain."
• Digital graphics and vector: "Flat vector illustration, clean lines, solid color blocking, minimalist graphic design."
• Classic painting: "Impressionist oil painting, textured impasto brushstrokes, layered canvas finish, warm palette."
• 3D rendering: "3D architectural render, Octane Render, ray-traced reflections, subsurface scattering."
• Anime and comics: "Cel-shaded anime key visual, dynamic action lines, crisp ink contours, vibrant color fill."
• Pixel and retro: "Pixel art, 16-bit graphics, blocky low-resolution sprite, limited palette."

Before committing a style to a paid campaign, verify the licensing position for that engine in our guidance on commercial usage rights for generated artwork and the style-accuracy comparison in our best AI art generator review.

Quick-Start Modular Templates (Fill-in-the-Blank)

Copy a template, replace each bracketed slot, delete what you do not need:

• Photorealism template:

• Editorial / blog photo template:

  Editorial photo, [Subject with age and appearance detail], [Action verb, e.g., reviewing a contract], [Expression, e.g., focused expression], [Setting, e.g., sunlit home office], [Prop or accent detail]

• Vector / branding template:

  Flat vector logo of [Subject], [Style, e.g., minimalist geometric], [Color Palette, e.g., dual-tone navy and gold], isolated on solid white background, vector graphic, no gradients, clean lines, no text

• 3D render template:

  3D architectural render of [Subject/Building], [Material Specs, e.g., frosted glass and polished steel], illuminated by [Lighting, e.g., volumetric sunset light], Octane Render, ray-traced reflections, detailed material textures

• Cinematic template:

  Cinematic [Shot Scale] of [Subject/Action], [Environment at Time of Day], [Atmospheric Effect, e.g., volumetric fog], [Lighting Direction, e.g., hard side light], 35mm film still, [Color Grade] --ar 2.39:1

• Illustration / anime template:

To compare performance across creative workflows, consult our AI Media Comparison overview.

To compare performance across creative workflows, consult our AI Media Comparison overview.

Match Prompt Detail to the AI Model

Format text prompts to match the natural language understanding and token parsing style of the chosen engine. Architecture decides whether long prose or compact keyword strings win.

• DALL·E 3: Optimised for natural language prose. It internally rewrites brief prompts into detailed descriptions, so descriptive conversational sentences beat isolated tag lists (OpenAI Prompt Engineering Guide, 2026).
• Midjourney: Responds best to compact, phrase-based syntax separated by commas. It leans heavily on appended parameter flags such as --ar 16:9 or --stylize 250 for execution control (Midjourney Documentation, 2026). Weighing this engine against alternatives? See our analysis of Midjourney image generation versus competing tools.
• Stable Diffusion and Flux: Require precise keyword weighting, explicit positive and negative prompt fields, plus technical parameters like CFG scale and sampler selection to control latent denoising (Stability AI Docs, 2026). Flux.1 additionally rewards concise scene descriptions and reference-image conditioning over very long token lists.
• Adobe Firefly 3 to 5: Prefers clear, context-first everyday language, then incremental refinement by adding style and material terms. Commercial-safety filters actively reshape style requests, so expect substitutions.

Meta-Prompting: How to Use LLMs to Write Image Prompts

Large language models make excellent prompt compilers. They expand a rough concept into a syntactically correct, engine-specific string much faster than manual drafting. The trick is constraining the output format so the LLM returns a prompt, not an essay about your prompt.

System Prompt for ChatGPT/Claude to generate Midjourney Prompts:
"You are an expert AI Image Prompt Engineer. I will give you a rough image concept.
Your task is to expand it into a precise, comma-separated Midjourney V6 prompt.
Structure your output as follows:
[Main Subject & Action], [Environment & Background], [Camera Lens & Framing], [Lighting Setup], [Artistic Medium/Style], [Color Palette] --ar 16:9 --v 6.0 --stylize 250.
Do not use filler words like 'hyperrealistic' or 'generate an image of'.
My concept is: [INSERT YOUR CONCEPT HERE]"

Three variations of the same meta-prompt cover most production needs:

1. Prose variant (DALL·E 3 / Firefly)replace the structure line with "Return one descriptive sentence of 40 to 60 words covering subject, setting, lighting, lens and one explicit exclusion. Use no parameter flags."
2. Dual-field variant (Stable Diffusion / Flux)add "Return three labelled blocks: POSITIVE, NEGATIVE, SETTINGS (CFG, sampler, steps, seed, resolution)."
3. Batch variant (A/B testing)add "Return five numbered prompt variants that differ in exactly one element, the lighting descriptor, keeping all other tokens identical."

Always review LLM-written prompts before execution. Language models cheerfully insert banned evaluative words, living artists' names or brand references that break both prompt adherence and compliance policy.

Use Tool Settings Alongside the Text Prompt

Combine textual description with native generator settings (aspect ratio, prompt weight or CFG scale, random seed) to control generation structure. Adjusting platform parameters stops you overloading the text prompt with instructions the system flags handle better.

Setting aspect ratio via --ar 16:9 in Midjourney, or explicit width and height in Stable Diffusion, secures framing without spending composition tags. The Classifier-Free Guidance (CFG) scale decides how strictly the model adheres to prompt tokens versus exploring the latent space, and values between 6.0 and 9.0 give solid prompt adherence without visual burn (Stability AI Docs, 2026). Midjourney's --stylize runs on a 0 to 1000 scale with a default of 100: low values track the text closely, high values let the model impose its own aesthetic. Sampler choice (DPM++ 2M Karras, Euler a, DDIM) changes both convergence speed and micro-texture, so fix it before any prompt A/B test. Otherwise you are measuring two variables and reporting one.

How to Write Prompts for Image-to-Image (Img2Img) and Inpainting

Text-to-image is only half of production work. Image-to-image, inpainting and outpainting start from an existing asset (a photo, a sketch, a rendered layout) and use the prompt to describe the delta, not the whole scene again.

The controlling variable is image strength, sometimes labelled denoising strength: a 0.0 to 1.0 value deciding how much of the source latent survives.

0.15 to 0.30colour and texture polish only; composition and identity preserved.

0.30 to 0.50material, wardrobe and lighting changes; geometry preserved. Best default for inpainting.

0.50 to 0.75style transfer and sketch-to-render conversion; composition loosely preserved.

0.75 to 1.0effectively a new generation, loosely guided by the source.

Img2Img prompting formula:

[Source image] + [Masked region / inpaint segment] + [Prompt describing only the change] + [Denoising strength: 0.3 to 0.75]

Worked example, change a garment colour and fabric on a finished portrait:

• Wrong (re-describes the whole scene, so the model regenerates the person): "A portrait of a woman in a red sweater standing in a park."
• Right (mask applied to the sweater only): Red cashmere knit sweater, detailed fabric weave, soft natural folds --denoise 0.45

Worked example, sketch to product render:

• Source: pencil sketch of a bottle. Prompt: matte ceramic bottle, soft studio softbox lighting, neutral seamless backdrop, product photography, subtle contact shadow · Denoise: 0.65 · Negative: text, label, watermark, harsh reflections.

Worked example, extending a frame (outpainting): keep the original prompt tokens for style continuity and add only the new spatial content: same editorial lighting, continued marble floor, blurred office windows on the left. Our comparison of AI outpainting tools shows how far each engine can extend a canvas before texture drift appears.

Three structural conditioning options deserve a place in any pipeline:

1. ControlNet: pass an edge map, depth map or OpenPose skeleton alongside the text prompt to lock geometry while the prompt controls surface appearance. Multiple conditionings can be combined in a single generation (Adding Conditional Control to Text-to-Image Diffusion Models, Zhang & Agrawala, 2023).
2. Reference or style conditioning: Midjourney --sref and --cref, Flux reference inputs and Firefly style references transfer palette and rendering treatment without copying composition.
3. Prompt-to-prompt editing: swap a single word in the prompt while preserving cross-attention from the unchanged tokens, which localises the edit without re-rolling the layout (Prompt-to-Prompt Image Editing with Cross Attention Control, Hertz et al., 2022).

For tool performance across standardised prompt sets, review our latest benchmarks.

Photorealistic, Cinematic and Portrait Prompts

Photorealistic portraits and cinematic scenes need explicit camera lenses, studio lighting setups and realistic texture modifiers. These prompts emulate real-world photographic equipment and environmental conditions. Corpus analysis explains why the same three ingredients recur in every strong example: camera description, lighting type and realism tags form their own thematic clusters in successful portrait prompts.

• Studio executive portrait:

  "A formal studio portrait of a corporate executive, three-quarter angle view, neutral expression, subtle skin texture. Key light softbox setup with gentle fill light, shallow depth of field, neutral gray background. Shot on 85mm f/1.8 lens, sharp focus on eyes, muted color palette."

• Cinematic environment scene:

  "A cinematic wide-angle shot of a modern financial trading floor at night, empty glass desks, glowing monitor screens reflecting blue and amber light. Atmospheric volumetric fog, dramatic side lighting, high contrast shadows, 35mm film still aesthetics, widescreen aspect ratio."

• Documentary workplace portrait:

  "Editorial photo, compliance officer in her mid-fifties, greying hair in a low knot, navy blazer, reviewing printed reports, focused expression, glass-walled meeting room, overcast window light, 50mm lens."

Teams standardising headshots across an organisation should also review capability limits in our guide to AI headshot generators, then finish frames with the retouching workflows in our overview of AI photo editors.

Illustration, Painting and Anime Prompts

Prompts for Design, Branding and 3D Images

Test Variations Instead of Rewriting Everything

Isolate individual variables when testing prompt performance instead of rewriting the whole string between runs. Iterative parameter testing shows the visual impact of a single adjective or setting change with some precision.

A workable testing framework changes one element only, say the lighting descriptor, while subject, composition and seed stay fixed. Locking the random seed ensures differences between iterations come from prompt edits, not from random latent initialisation.

Reuse expectations should be calibrated per engine, because prompt-driven variability differs sharply between architectures.

Change One Prompt Element at a Time

Isolate visual defects by modifying a single phrase or parameter per iteration while generation settings stay constant. Systematic adjustment identifies the precise word causing distortion or misalignment.

When output shows wrong lighting, change only the lighting descriptor and re-run with the same seed. Editing several adjectives at once obscures which change fixed the problem and lengthens total iteration time.

Automated prompt-optimisation systems formalise the same discipline: they enrich only the missing concepts and leave satisfied tokens untouched.

Diagnostic shortcuts by defect type:

Use Negative Prompts and Advanced Prompting Techniques

Ethics, Copyright and Living-Artist Modifier Rules

Prompt text is a legal artefact as much as a technical one. Two constraints apply to any commercially published asset.

1. Names of living artists.Avoid direct triggers of the form in the style of [living artist]. Firefly and Midjourney v6 increasingly filter or substitute such tokens, and the ethical objection from working artists is substantive. Replace the name with the technique it stands for: impasto brushstrokes, cubist geometry, flat vector with limited palette, high-contrast chiaroscuro.
2. Trademarks and logos.Naming brands (Nike shoes, Apple iPhone) in commercial prompts creates trademark exposure and can reproduce protected trade dress. Use material descriptors instead: minimalist aluminium smartphone, unbranded, white low-top leather sneaker, no logos.

Three further practices are worth institutionalising: avoid generating recognisable likenesses of real individuals without consent; avoid prompts that could produce misleading depictions of real events; and document the engine, model version and licence terms attached to every published asset.

Enterprise Risk, Brand Safety and Prompt Governance

In regulated industries, prompt text should be governed like any other input that leaves the perimeter. That framing usually settles the argument with marketing faster than a policy memo.

• No confidential data in prompts. Never paste client PII, account numbers, internal project code names, unreleased product specifications or sensitive strategy into public generation services. Prompts submitted to consumer tiers may be retained or used for service improvement, depending on the provider's terms.
• Prompt logging and audit trail. Store prompt text, negative prompt, seed, model version, parameter set, operator identity and timestamp for every published asset. That record is what makes an output reproducible, and defensible, months later.
• Shadow AI control. Unapproved tools produce assets with unknown licence provenance. Maintain an approved-engine list with documented commercial terms and a named owner for each entry.
• Brand-safety review. Route generated assets through the same approval workflow as photography: palette compliance, inclusivity review, and a check for inadvertent text, logos or protected marks.
• Documented assumptions and limits. Record intended use, assumptions and known limitations for each generative workflow, consistent with the documentation practices in the NIST Generative AI Profile (NIST AI 600-1, 2024).
• Verification of inbound assets. When receiving images from agencies or contributors, screening with AI image detectors helps confirm whether a declared photograph is in fact synthetic.

Following these best practices for AI image generation prompts measurably reduces rework. Run the pre-generation checklist below before anything reaches a production queue.