Conclusions
No Difference in Major Semantic Attributes
Across 3 scenes × 3 seeds = 9 pairs, composition, pose, clothing color, expression, lighting direction, and presence of key objects were reproduced at roughly equal quality in both formats.
Detail-Level Tendencies
| Tendency | Tag List | Natural Language |
|---|
| Control of individual elements | Slightly higher (cherry blossom density, background cleanliness) | — |
| Softness of lighting | — | Tends slightly softer |
| “Cozy” atmosphere | — | Contextual additions like cushioned benches |
| Wide vs close composition | Slightly wider | Slightly closer |
Practical Conclusion
Either is fine. There is no significant difference in key attribute reproduction — choose based on personal preference.
If you must choose:
- When you need precise element control → Tag list (each attribute tends to be processed as independent tokens)
- When prioritizing overall atmosphere → Natural language (contextual supplementation creates cohesion)
The Question
There are two writing styles for AI image generation prompts:
- Tag list:
1girl, red dress, standing, park, cherry blossoms - Natural language:
A girl in a red dress standing in a park with cherry blossoms
Which is better? Does it make a difference? We test with fixed seeds.
Experiment Conditions
| Parameter | Value |
|---|
| Model | z-image-turbo (6B, photorealistic distilled model) |
| Steps | 8 |
| Sampler | euler / ddim_uniform |
| CFG | 1.0 |
| Image Size | 1024×1024 |
| Seeds | 100, 200, 300 (fixed) |
For each of 3 scenes, we prepare a tag list version and a natural language version with the same meaning, then generate with identical seeds and compare.
Scene 1: Cherry Blossom Park
1girl, 22yo japanese woman, red dress, standing, park, cherry blossoms, spring, sunny day, smiling
A 22 year old Japanese woman in a red dress standing in a park with cherry blossoms on a sunny spring day, smiling.
seed=100
| Tag list | Natural language |
|---|
 |  |
seed=200
| Tag list | Natural language |
|---|
 |  |
seed=300
| Tag list | Natural language |
|---|
 |  |
Cherry Blossom Scene Observations
| Attribute | Difference |
|---|
| Dress color (red) | Equal. All 3 pairs are red |
| Expression (smiling) | Equal |
| Cherry blossom density | Tag list slightly richer in cherry blossom depiction (2/3 pairs) |
| Background cleanliness | Tag list has slightly cleaner park background (2/3 pairs) |
| Lighting | Natural language tends softer light (2/3 pairs) |
| Composition | Tag list tends slightly wider (2/3 pairs) |
Scene 2: Library
1girl, 22yo japanese woman, sitting on bench, reading book, library, wooden bookshelves, warm lamp light, glasses, focused expression
A 22 year old Japanese woman wearing glasses sits on a bench in a library with wooden bookshelves, reading a book under warm lamp light with a focused expression.
seed=100
| Tag list | Natural language |
|---|
 |  |
seed=200
| Tag list | Natural language |
|---|
 |  |
seed=300
| Tag list | Natural language |
|---|
 |  |
Library Scene Observations
| Attribute | Difference |
|---|
| Glasses presence | Equal. Both formats reproduced in all 3 pairs |
| Wooden bookshelves | Equal |
| Reading pose | Equal |
| Warm lamp light | Equal |
| Focused expression | Equal |
| Bench texture | Difference: Tag list 3/3 bare wood. Natural language 3/3 cushioned bench |
| Field of view | Natural language slightly wider (2/3 pairs) |
The reason is unclear, but natural language tended to produce cushioned benches.
Note: The natural language version showed a tendency to produce a cushioned bench, suggesting a form of context supplementation. This is speculative, but “sits on a bench in a library” may have been interpreted as “a comfortable place to sit.”
Scene 3: Rainy Night Street
1girl, 22yo japanese woman, black leather jacket, walking, rainy street at night, neon signs, wet pavement reflections, looking back over shoulder
A 22 year old Japanese woman in a black leather jacket walks down a rainy street at night with neon signs, wet pavement reflecting the lights, looking back over her shoulder.
seed=100
| Tag list | Natural language |
|---|
 |  |
seed=200
| Tag list | Natural language |
|---|
 |  |
seed=300
| Tag list | Natural language |
|---|
 |  |
Rainy Night Street Scene Observations
| Attribute | Difference |
|---|
| Composition | Equal. Nearly identical in all 3 pairs |
| Looking back pose | Equal |
| Leather jacket | Equal |
| Neon signs | Equal. Placement and density similar |
| Pavement reflection | Equal. Both formats are vivid |
| Overall mood | Equal |
There was essentially no difference in this scene. Differences between same-seed pairs were only in sign text details and minor outfit details — within seed variation range.
Note: The fact that major attributes don’t differ between tag list and natural language means either format can be used without a major risk of failure. For ease of adding and removing elements, the tag list format is the more practical choice.
Related Articles
![[Verified] Image Generation Prompt Best Practices](/tips/prompt-best-practices/cover_0_0000_4517457392071889496.webp)
