Writing Descriptive Prompts for AI Image Generation
Image models can't read your mind, only your text. Learn the protagonist + visual style + composition formula to go from an "average" image to the scene you imagine.
Image models can't read your mind, only your text. A vague prompt returns the model's "average" interpretation (clean, tidy, predictable). A descriptive prompt is the difference between a generic image and the exact scene you imagine. It's like describing a scene over the phone to someone who can't see it: in enough detail that they can reconstruct it without error.
Why generic prompts fail
"Generate an image of a modern office" gives a tidy, predictable, average office. But if you pictured a chaotic, creative open space (Google or Facebook style), the output falls short. Generic input leads to a random, "average" output.
The 3-part formula
Prompt = Protagonist + Visual Style + Composition1. Protagonist (the subject). Add concrete attributes instead of general words.
a big doga golden retriever with a red collar, medium height, about two years old, very playfulAttributes to specify: color, size, age, attitude or personality, and accessories. It's the difference between "bring me a glass from the kitchen" (random result) and "bring me the blue glass next to the microwave" (precise result). More concrete attributes means more control.
2. Visual style. Think of it as an Instagram filter applied to your scene: it changes the result completely. Examples: realistic phone photo, cartoon, Disney-style character, old grainy photo, modern corporate photography. It's like choosing an outfit for a wedding: "workout clothes" is vague; a formal suit with a bow tie, a specific shirt and exact colors is precise. Style gives the image its life and personality.
3. Composition (act as a director of photography). Define how the scene is shot:
- Camera angle: table level, 45 degrees, from the corner of the room, overhead, etc.
- Shot height: where the camera sits relative to the subject.
- Lighting: natural light through a window vs. intense white ceiling light.
- Atmosphere: environmental details such as smoke, fog or reflections that add depth.
Worked example
People working.Generate an image of four diverse professionals collaborating around a round table with laptops and documents, in a modern corporate photography style, with a 45-degree angle view and natural office light.| Part | In the example |
|---|---|
| Protagonist | four diverse professionals, round table, laptops, documents |
| Style | modern corporate photography |
| Composition | 45-degree angle, natural office light |
Common mistake: contradictory instructions
Combining attributes that cancel each other out confuses the model and triggers hallucinations. Example: "a realistic photo in cartoon style". It's like asking someone to sing at full volume but quietly.
The fix: choose between opposing styles first, then keep the protagonist, style and composition consistent with each other.
When the result is wrong
Treat it like cooking: try, taste, adjust.
- Give specific feedback about what didn't work.
- Adjust the prompt and regenerate.
- Ask the model what the ideal prompt would have been to produce that result.
Template and checklist
Generate an image of [protagonist with specific attributes], in [visual style], with [camera angle/height] and [lighting/atmosphere].- Does the protagonist have concrete attributes (color, size, age, attitude, accessories)?
- Is the visual style explicit (photo, cartoon, Disney, vintage, etc.)?
- Are camera angle, height and lighting defined?
- Did I add atmosphere details (smoke, fog, reflections)?
- Are there any contradictions between style, subject and composition?
Practice: generate an image with ChatGPT of the scene you are in right now (taking the class), using the three-part formula, and share the image together with the prompt you used.
Key takeaways
- A generic prompt returns the average result; specificity gives you control.
- Protagonist + visual style + composition.
- Avoid contradictory instructions between style, subject and composition.
- Iterate like cooking: try, give specific feedback and adjust.