At Wolseley Fine Arts, we work at the intersection of traditional craftsmanship and modern creative tools. AI image generators have become part of every designer’s toolkit, but the gap between a mediocre output and a stunning visual almost always comes down to one thing: the prompt. This guide is a hands-on walkthrough of how to write prompts for AI image generation, with side-by-side comparisons of weak and strong prompts so you can see exactly why some prompts produce magic and others produce mush.
Why Prompt Writing Is a Designer’s New Core Skill
A well-crafted prompt is essentially a design brief written for a machine. It must describe the subject, the composition, the mood, the lighting, the medium, and the style, all in a way the model can parse. Tools like Midjourney v7, Stable Diffusion 3.5, and Adobe Firefly all reward structured, specific language and punish vague instructions.
The Anatomy of a Strong AI Image Prompt
Every effective prompt, regardless of the platform, follows a consistent skeleton. Here is the structure we use internally at the studio:
- Subject: the main noun (what is being depicted)
- Action or pose: what the subject is doing
- Environment: where the scene takes place
- Composition: framing, angle, shot type
- Lighting: quality and direction of light
- Style or medium: oil painting, photography, 3D render, etc.
- Artist or reference: optional stylistic anchor
- Technical modifiers: resolution, lens, aspect ratio
Image generation prompts typically work best when written as short phrases separated by commas rather than full grammatical sentences. The model parses tokens, not prose.
Weak vs Strong Prompts: Four Practical Comparisons
Example 1: A Portrait
| Weak Prompt | Strong Prompt |
|---|---|
| a woman in a garden | portrait of a woman in her 40s, soft smile, standing in an English cottage garden, roses and foxgloves, golden hour backlight, shallow depth of field, shot on Hasselblad 80mm, editorial photography, photorealistic, 4k |
Why the strong version wins: the weak prompt gives the model no anchor for age, mood, framing, lighting, or medium. The strong version locks down the photographic medium, the lens, the time of day, and even the specific flowers, producing a controlled, gallery-ready image instead of a generic stock photo.
Example 2: A Product Visual
| Weak Prompt | Strong Prompt |
|---|---|
| a perfume bottle on a table | amber glass perfume bottle, gold cap, centered on a polished travertine surface, dried lavender stems beside it, soft diffused window light from the left, minimalist luxury still life, beige and cream palette, 3:4 aspect ratio, hyper detailed |
Why it works: e-commerce and editorial product shots need controlled lighting and palette. Naming the surface material, the secondary props, and the light direction transforms a flat render into a usable brand asset.
Example 3: A Fine Art Illustration
| Weak Prompt | Strong Prompt |
|---|---|
| a fantasy forest | ancient oak forest at twilight, moss covered roots, beams of indigo light filtering through canopy, mist near the ground, painterly illustration, gouache and ink, inspired by Arthur Rackham and John Bauer, intricate linework, muted jewel tones |
Why it works: referencing artists whose work the model has likely been trained on (when terms of use allow) gives the generator a clear stylistic target. Specifying medium (gouache and ink) prevents the default photorealistic drift.
Example 4: An Architectural Concept
| Weak Prompt | Strong Prompt |
|---|---|
| modern house in the mountains | contemporary cantilevered house, charred timber cladding, floor to ceiling glazing, perched on a rocky alpine ridge, low morning fog, wide angle exterior architectural photograph, 24mm lens, overcast diffused light, brutalist minimalism, ArchDaily editorial style |
Modifiers That Consistently Improve Output
These are the modifier categories we lean on most. Mix and match, but avoid stacking contradictions.
- Lighting: golden hour, rim lighting, chiaroscuro, soft box, volumetric light, neon glow
- Camera and lens: 35mm, 85mm portrait lens, macro, drone shot, fisheye, tilt shift
- Composition: rule of thirds, symmetrical, centered, overhead shot, dutch angle, close up
- Medium: oil painting, watercolor, charcoal sketch, 3D octane render, isometric, vector illustration
- Mood: serene, melancholic, dramatic, nostalgic, ethereal
- Detail and quality tags: hyper detailed, intricate, photorealistic, sharp focus, 8k
Platform Specific Tips
Midjourney
- Use parameters at the end: –ar 16:9 for aspect ratio, –stylize to control artistic interpretation, –chaos for variation
- Reference images using image URLs at the start of the prompt for style transfer
- Keep prompts under roughly 60 words; Midjourney rewards conciseness
Stable Diffusion
- Use weighted terms with parentheses: (golden hour:1.3) increases that token’s influence
- Always pair with a negative prompt to exclude unwanted elements (blurry, deformed hands, extra fingers, watermark, low quality)
- LoRAs and custom checkpoints often outperform prompt engineering alone for niche styles
Adobe Firefly and DALL-E
- Both respond well to natural language mixed with comma separated descriptors
- Firefly favors commercial safe styles, so be explicit about medium and palette
Common Mistakes to Avoid
- Vague subjects: “a cool character” tells the model nothing
- Overloading the prompt: more than 10 to 12 distinct concepts dilutes each one
- Contradictory styles: “photorealistic watercolor cartoon” produces visual noise
- Ignoring aspect ratio: portrait, landscape, and square framing dramatically change composition
- Forgetting negative prompts on Stable Diffusion based tools
A Reusable Prompt Template
Save this template and adapt it for your next project:
[Subject], [action or pose], [environment], [composition and framing], [lighting], [medium or style], [artist or reference], [mood], [technical modifiers] –ar [ratio] –stylize [value]
FAQ
How long should an AI image prompt be?
Aim for between 15 and 50 words. Microsoft Copilot recommends a minimum of six descriptive keywords, and in our testing the sweet spot for most platforms sits around 25 to 35 words of dense, comma separated descriptors.
Should I write in full sentences or keywords?
Short phrases separated by commas almost always outperform full sentences for image models. Save full sentences for narrative or dialogue work with text based AI.
Can I reference living artists in my prompt?
Technically yes on many platforms, but ethically and legally it is increasingly questionable. We recommend referencing artistic movements, historical artists in the public domain, or descriptive style language instead.
Why do my prompts produce different results each time?
AI image generators use random seeds. To reproduce a result, fix the seed value in tools that allow it (Stable Diffusion, Midjourney via parameters). This is essential for iterating on a single direction.
What is a negative prompt?
A negative prompt lists what you do not want in the image. It is most useful in Stable Diffusion based tools and helps eliminate common artifacts like distorted hands, extra limbs, blurry backgrounds, and unwanted text.
Final Thoughts
Learning how to write prompts for AI image generation is less about secret magic words and more about thinking like an art director. Define the subject, set the scene, control the light, choose the medium, and refine through iteration. The prompt is the brief, and the better your brief, the better your output. At Wolseley Fine Arts we treat every prompt the same way we treat a commissioned piece: with intent, reference, and craft.