Image Generation
Image generation prompts (overview)
Image generation prompts (overview)
A reusable prompt, example set, evaluation record or safety rule with its task boundary preserved.
Run at least one representative case, inspect the result and record what still requires human review.
This section collects prompts for exploring image generation / multimodal capabilities. The focus is on iterative prompt refinement to gradually steer the output toward what you want. The core of image tasks: clear visual goal + structured description + repeatable iteration strategy.
Common Scenarios
- Product visual direction exploration (style, mood, color palette)
- Quick visual asset drafts (posters, covers, illustrations)
- Brand identity generation (characters, mascots, IP)
- Scenario concept validation (interior/street/futuristic)
Key Elements (spell these out every time)
- Subject: Person/object/scene
- Style: Realistic/illustration/pixel/3D/paper craft
- Composition: Angle, depth of field, subject position
- Lighting & color tone: Warm/cool, soft/hard, time of day
- Detail constraints: Materials, textures, key elements
Prompt Template (General)
Subject: {{SUBJECT}}
Style: {{STYLE}}
Composition: {{COMPOSITION}}
Lighting: {{LIGHTING}}
Color tone: {{COLOR_TONE}}
Details: {{DETAILS}}
Example 1: Brand Visual Direction
Subject: Hero poster for an AI learning brand
Style: Modern flat illustration, slight paper texture
Composition: Center-symmetric, main title centered, whitespace below
Lighting: Soft light, subtle gradient
Color tone: Warm palette, orange + off-white + dark blue accents
Details: Abstract circuit lines, learning elements (books, notes)
Example 2: Product Scene
Subject: Smart learning device on a desk
Style: Realistic photography
Composition: 45-degree overhead angle, shallow depth of field
Lighting: Afternoon natural light, entering from the left
Color tone: Neutral, slightly warm
Details: Wooden desk, coffee cup, sticky notes
Example 3: Character / IP
Subject: A cute AI assistant cat
Style: 3D cartoon
Composition: Front-facing half body, smiling
Lighting: Soft studio lighting
Color tone: Blue and white color scheme
Details: Wearing small headphones, simple logo on chest
Iteration Strategy
- Rough first, details later: Lock down subject and style first, then add details
- Positive first, then constraints: Describe "what you want" first, then add "what you don't want"
- Change one variable at a time: Only tweak one element per iteration to isolate the effect
- Log versions: Save each prompt for easy rollback and comparison
Common Problems & Fixes
- Blurry subject: Add subject position and action descriptions
- Style drift: Pin down style keywords + reference objects
- Messy composition: Specify camera angle, depth of field, subject placement
- Missing details: Add material and key element qualifiers
Index
📚 Related resources
❓ Common questions
Open a question to review the practical answer.
Which elements should an image-generation prompt include?
The chapter lists five core slots: subject (person / object / scene), style (photo / illustration / pixel / 3D), composition (angle, depth of field, subject position), lighting (warm or cool, soft or hard, time of day), colour palette (main plus accent), and finally detail constraints (materials, key elements). Drop any slot and the model improvises; fill them all and the output stays on-style.
When an image prompt drifts, what is the most efficient fix?
The chapter gives three iteration rules: coarse before fine (lock subject and style first, add detail later), positive before negative (say what you want before what you do not), and change one variable per iteration. The common mistake is rewriting five words at once, then not knowing which one moved the result. Keep every prompt version as a changelog so you can roll back.
Can I reuse the same prompt template across Midjourney, Stable Diffusion and DALL-E?
The skeleton transfers, the keywords do not. The chapter's `subject / style / composition / lighting / colour / detail` skeleton works on all three platforms, but weighting syntax (Midjourney's `::`, SD's `(word:1.3)`) and negative prompts (only SD has a real negative-prompt field) do not. Reuse the skeleton across platforms and tune the details against each tool's docs.
If the subject keeps coming out blurry or off-position, how do I fix the prompt?
Three moves: (1) add subject position and action — e.g. `centred composition, headline middle, whitespace below`; (2) specify camera angle and depth of field — e.g. `45-degree side overhead, shallow DoF`; (3) lock style with named references — e.g. `modern flat illustration with subtle paper texture`. The brand-poster example in the chapter shows exactly that fully-specified structure.
For brand visuals, how do I keep multiple images on the same style?
Pin style, palette and composition as a locked block at the top of every prompt and only swap subject and detail below. Add a style-anchor phrase — e.g. `modern flat illustration, slight paper texture, warm palette: orange + cream + deep blue accent` — into every prompt. If the tool supports a reference image, feeding the same reference is the most stable way to enforce consistency.