AI image generation models respond very differently to how you phrase prompts. The gap between a mediocre AI image and a professional-quality one is almost always prompting technique, not model capability. This guide covers what works across DALL-E 3, Midjourney, and Flux.
The Core Principles
Be Specific About What You Want, Not What You Don't Want
Negative prompts ("no blurry background", "not cartoonish") work inconsistently. It's more effective to specify what you do want ("sharp background detail", "photorealistic style"). Midjourney supports negative prompts via --no, but positive specification is usually more reliable.
Lead with the Subject
Image models pay more attention to early tokens. Put your primary subject at the start of the prompt, not the end. "A golden retriever sitting on a red couch, warm afternoon light, indoor photography" works better than "warm afternoon light, indoor photography, red couch with a golden retriever."
Separate Subject from Style
Think of prompts as having two parts: the content (what) and the style (how). Clearly describing both leads to better results than mixing them.
Content: "a woman in her 40s reading a book at a cafe table"
Style: "warm natural light, candid photography, shallow depth of field, film grain"
Combined: "a woman in her 40s reading a book at a cafe table, warm natural light, candid photography, shallow depth of field, film grain"
Style Reference Language
These terms are understood by most major image models and reliably influence style:
Photography Styles
- Lighting: golden hour, blue hour, overcast diffused light, harsh midday sun, studio lighting, Rembrandt lighting, backlit
- Camera: 85mm portrait lens, wide angle, fisheye, macro photography, aerial shot, eye-level shot
- Film qualities: film grain, Kodak Portra 400, cinematic, analog photography, lomography
- Depth of field: shallow depth of field, bokeh background, everything in focus, tilt-shift
Art Styles
- Mediums: oil painting, watercolor, ink drawing, pencil sketch, charcoal, gouache, digital art
- Movements: impressionist, art nouveau, art deco, baroque, surrealist, minimalist
- Rendering styles: concept art, illustration, graphic novel, anime, Studio Ghibli style, pixel art, isometric
Quality Modifiers
Some quality-indicating terms consistently improve output quality in most models:
- highly detailed, intricate details
- professional photography
- 8K resolution (even if not literally 8K, signals quality)
- award-winning photography
- ultra-realistic
--q 2 for higher quality in Midjourney instead.
Model-Specific Prompt Techniques
DALL-E 3
DALL-E 3 is the most natural language-friendly of the three. You can write in complete sentences and it will follow them reliably.
- Write in descriptive prose, not keyword lists
- Be specific about spatial relationships: "in front of", "to the left of", "behind"
- Explicitly request text content: "a storefront sign that reads 'Grand Opening'"
- When using ChatGPT, ask it to refine prompts: "Write me a DALL-E 3 prompt for..."
Example: "A cozy independent bookshop interior, late afternoon sunlight streaming through large windows, wooden shelves floor to ceiling filled with colorful books, a tabby cat sleeping on a reading chair in the foreground, warm amber tones, soft focus background, inviting atmosphere"
Midjourney
Midjourney performs best with keyword-dense prompts and style references. The Discord interface supports special parameters:
- Use
--arto set aspect ratio:--ar 16:9for widescreen,--ar 9:16for portrait - Use
--style rawfor more photorealistic, less "Midjourney aesthetic" results - Use
--stylize(0–1000) to control how strongly Midjourney applies its aesthetic preferences - Reference other images with
--sreffor style consistency - Use
--nofor negative prompts:--no text, watermark
Example: portrait of a weathered sailor in his 60s, North Sea, overcast light, documentary photography, Canon 5D, 85mm, f/2.8, Kodak Portra 400, intricate detail --ar 4:5 --style raw
Flux
Flux handles both prose and keyword approaches well. It's more sensitive than Midjourney to explicit style instruction and less opinionated about its own aesthetic preferences.
- Works well with both detailed prose and keyword lists
- Explicit style language is effective: "cinematic composition", "editorial photography"
- For Flux Schnell (fast model), simpler prompts often work better than complex ones
- For Flux Dev/Pro, detailed prompts get more detailed results
Prompt Formulas That Work
Photography Formula
[subject] + [context/environment] + [lighting description] + [camera/lens] + [mood/aesthetic]
Example: "young architect reviewing blueprints + modern office, large windows, Tokyo skyline in background + diffused morning light + 50mm lens, shallow depth of field + focused, professional, quiet atmosphere"
Art/Illustration Formula
[art medium] + [subject and scene] + [color palette] + [artistic movement/style]
Example: "oil painting + fishing village at sunset, small wooden boats, seagulls + muted blues and warm oranges + impressionist, Monet-inspired brushstrokes"
Product/Commercial Formula
[product description] + [setting] + [lighting] + [photography style] + [quality indicators]
Example: "glass perfume bottle with gold cap + white marble surface, flower petals scattered around + soft studio lighting, subtle shadows + commercial product photography + clean background, highly detailed"
Common Mistakes to Avoid
| Mistake | Problem | Fix |
|---|---|---|
| Too vague | "a beautiful landscape" | Specify season, location, time of day, weather, mood |
| Too many subjects | 5+ people doing different things | Focus on 1–2 main subjects |
| Conflicting styles | "photorealistic cartoon anime" | Pick one consistent aesthetic |
| Relying only on negative prompts | "no ugly faces" | Specify what you want: "attractive, natural expression" |
| Expecting exact text | Long sentences in images | Keep to 1–3 words, use DALL-E 3 specifically |
Iterative Refinement
Professional image generation workflows treat prompting as iterative:
- Start with the core subject to verify the model's interpretation
- Add environment/context once the subject looks right
- Add style, lighting, and quality modifiers last
- When a result is close but not right, adjust one variable at a time to understand what changes
- Save prompts that work as templates for similar future images
Frequently Asked Questions
How long should my prompt be?
For most models, 50–150 words is the sweet spot. Very short prompts leave too much to chance. Very long prompts (300+ words) can confuse the model or cause later instructions to be ignored. Midjourney tends to work well with shorter, keyword-dense prompts; DALL-E 3 and Flux handle longer prose well.
Why do the same prompts produce different results each time?
Image generation models are probabilistic — the same prompt produces different images each run. Use seed parameters when available to reproduce specific results. Midjourney's "make variations" feature helps iterate from a specific result you like.
Can I prompt for a specific person's face?
Prompting for specific real people's faces raises significant legal and ethical issues around likeness rights, deepfakes, and consent. Most models have policies against this. Midjourney and DALL-E 3 actively filter for celebrity likenesses. Use general descriptors (age, hair color, expression) rather than named individuals.
