21.【Generative AI Experiment】How Much “Kyoto” Can Change
Comparing 10 Prompt Directive Variations in Image Generation
tags:
- GenerativeAI
- ImageGeneration
- Prompt
- Experiment
- Kyoto
🎯 Introduction
When using image-generation AI, it’s easy to fall into this situation:
“The image looks plausible, but I don’t know why it turned out that way.”
In this article,
we fix the theme to “Kyoto”,
and change only the “directive words” inside the prompt to observe:
- What changes
- Which visual elements change
- And how much they change
🎯 Goal
👉 To understand prompts
not as “magic spells,” but as control inputs.
🧠 Why Choose “Kyoto”?
“Kyoto” is an excellent subject for observing image-generation behavior.
- 🏙 A real city with abundant training data
- 🏯 A mix of tradition, tourism, daily life, and modern urban space
- 🎨 Easily shifts between realism, illustration, and abstraction
In short:
Small wording differences produce visually obvious changes
This article is
not about finding the “correct” image, but about observing change.
🔬 Experimental Conditions (Fixed Rules)
The following conditions are fixed throughout the experiment:
- Theme word: Kyoto / 京都
- Variables per run: only one changed element
- Models: DALL·E / Stable Diffusion / Midjourney, etc. (not specified)
- Observation targets:
- Composition
- Atmosphere
- Realism
- Abstraction level
🔬 Ensure that “what was changed” is always unambiguous.
⚙️ Base Prompt (Common to All Tests)
The following prompt is used as the common base,
and only the differences are modified in each experiment.
Kyoto landscape,
clean composition,
natural lighting,
no text, no logo, no watermark
⚙️ The goal here is not to generate a “good” image,
but to make differences easy to observe.
🔬 “Kyoto” Fixed — 10-Pattern Experiment Table
| No. | Change Type | Prompt Difference | Main Visual Change | Observation Focus |
|---|---|---|---|---|
| 1 | Noun | Kyoto landscape photograph |
Photo / tourism-like | Attraction to real locations |
| 2 | Noun | Kyoto landscape illustration |
Poster / artwork style | Color and composition cleanup |
| 3 | Abstraction | abstract visualization of Kyoto landscape |
Buildings dissolve | Atmosphere over meaning |
| 4 | Specificity | realistic Kyoto street scene |
Everyday life, alleys | Increased realism |
| 5 | Emotion | calm and quiet Kyoto landscape |
Static, low saturation | Light and color suppression |
| 6 | Emotion | dramatic Kyoto landscape |
Overstaged visuals | Strong clouds and sunsets |
| 7 | Era | Kyoto scenery in Edo period style |
Nihonga / woodblock style | Style overrides realism |
| 8 | Era | modern Kyoto cityscape |
Contemporary urban feel | Deviation from “tourist Kyoto” |
| 9 | Viewpoint | aerial view of Kyoto |
Bird’s-eye composition | Emphasis on geography |
| 10 | Viewpoint | street-level view of Kyoto |
Human-eye perspective | Depth and narrative |
🔬 Generated Image Comparison (10 Patterns)
Figure 1: Ten generated results with “Kyoto” fixed,
changing only directive words in the prompt.
🧠 Observations and Findings
① The First Noun Defines the Worldview
photographillustrationvisualization
This single noun at the beginning
almost completely determines the image’s direction.
🧠 The most important word for locking the visual world.
② Abstract Words Dissolve “Meaning”
Adding words like:
abstractdreamlikeinspired by
causes intentional ambiguity in:
- Building shapes
- Road structures
- Spatial consistency
👉 Ambiguity is not a bug, but an exploration parameter.
③ Emotion Words Control Color and Light
- calm / quiet
→ lower saturation, morning/evening light, mist - dramatic
→ strong contrast, backlighting, dramatic clouds
🧠 Emotion words do not control emotions,
they control visual staging.
⚠️ ④ Era Specification Is the Strongest Parameter
Words like:
Edo periodmodern
instantly override:
architecture, clothing, color palette, and composition
⚠️ Extremely powerful, but often ignores other instructions.
⑤ Viewpoint Fixes the Composition
aerial viewstreet-level view
almost forcibly determine:
- Camera position
- Depth
- Information density
👉 If composition control is the goal, this should be prioritized.
⚙️ Practical Implications
- 🎯 Eyecatch images
→ Start with noun + emotion - 📘 Educational diagrams
→ Explicitly usediagram / schematic / precise - 🎨 Atmospheric visuals
→ Abstract terms + negative prompts - 🔬 Comparative experiments
→ Fixed theme + one-word difference
✅ Summary
- Prompts are not magic spells
- Prompts are control inputs
- Ambiguity is not bad — it is data
- Changing just one word can dramatically alter results
Start with:
“Fixed theme × one-word difference”
This is the fastest way to truly understand prompt behavior.
End.