Conclusion
- Compressing from 200+ tokens to ~40 tokens showed no degradation in image quality or core element reproduction
- Camera model names, lens specs, aperture, ISO, and other shooting parameters have no effect on output
- Quality keywords like
ultra-detailed skin textureandsharp focusare also ineffective - Repeating the same concept three times is no better than writing it once — it just wastes tokens
Do you believe longer prompts produce better images? That writing camera model names and F-stops makes photos more realistic? We put that urban legend to the test.
The Prompt Under Test
The subject is a studio portrait prompt: a nude woman against a red gradient backdrop with a circular light.
This prompt has several issues.
Problem Analysis
1. Elements Verified as Ineffective
| Element | Category | Source |
|---|---|---|
Camera: Medium format, Hasselblad X2D / Sony A1 | Camera model | Confirmed ineffective in Bikini Prompt Iteration |
Lens: 85mm prime, Aperture: f/4 | Lens & aperture | Same as above |
Shutter Speed: 1/160, ISO: 100 | Shooting params | Same as above |
ultra-detailed skin texture | Quality keyword | natural skin texture confirmed ineffective in Prompt Optimization 10 Themes |
sharp focus | Quality keyword | z-image-turbo outputs sharp images by default |
2. Redundant Descriptions
The same concept appears in multiple places.
| Concept | Occurrences | Needed |
|---|---|---|
| Confident pose | standing confidently / confident stance / Calm dominant self-assured | Once is enough |
| Hand in pocket | one hand in his pocket / One hand in pocket | Once is enough |
| Dramatic lighting | dramatic lighting / Key Light, Fill Light, Rim Light details | One phrase |
3. Token Count Issues
CLIP processes 75 tokens per chunk. Influence drops off in subsequent chunks. At 200+ tokens spanning 3+ chunks, the camera specs and lighting details in the back half are likely being ignored entirely.
Optimized Version
Keeping only the core elements:
What Was Removed
- All camera/lens/shooting parameters — Camera, Lens, Aperture, Shutter Speed, ISO, White Balance, Focus
- Quality keywords —
sharp focus,ultra-detailed skin texture,premium magazine aesthetic,clean composition - Redundant expressions —
confident stance,Calm dominant self-assured,Slight lean, etc. - Detailed lighting breakdown — Key/Fill/Rim Light specs → consolidated into
dramatic front-left softbox lighting
What Was Kept
- Style —
A sophisticated studio portrait(sets overall direction at the front) - Subject —
1girl, 32yo japanese actress, full nude - Pose —
standing with one hand on hip(mentioned once) - Background —
bold red gradient backdrop, large glowing circular light behind her - Lighting —
dramatic front-left softbox lighting, high contrast - Finish —
luxury editorial style, cinematic color grading, modern minimalism - Composition —
3/4 body portrait, subject slightly off-center
Method
| Parameter | Value |
|---|---|
| Model | z-image-turbo |
| Steps | 8 |
| Sampler | euler |
| Scheduler | ddim_uniform |
| CFG | 1.0 |
| Image size | 1024×1024 |
| Seeds | 42, 123, 456 (3 images per condition) |
Note: Because the token sequences differ between prompts, images differ even with the same seed. This is the same phenomenon documented in Prompt Fundamentals regarding weight syntax — a side effect of token sequence changes, not a quality difference. The comparison target is whether core elements (red backdrop, circular light, nude, studio portrait) are equally reproduced.
Comparison Results
Seed 42
| Original (200+ tokens) | Optimized (~40 tokens) |
|---|---|
 |  |
Both reproduce the red gradient backdrop, circular light, and studio portrait style. No visible difference in skin texture or lighting quality.
Seed 123
| Original (200+ tokens) | Optimized (~40 tokens) |
|---|---|
 |  |
The original is more frontal while the optimized version has an angled composition — this is due to token sequence randomization, not quality difference. Red backdrop and circular light are reproduced in both.
Seed 456
| Original (200+ tokens) | Optimized (~40 tokens) |
|---|---|
 |  |
Stable composition in both. Skin texture is equivalent despite removing ultra-detailed skin texture from the optimized version.
Analysis
Core Element Reproduction
All core elements were reproduced across all 6 images (3 seeds × 2 conditions):
- Red gradient backdrop: 6 of 6
- Circular light: 6 of 6
- Studio portrait style: 6 of 6
- Nude: 6 of 6
No evidence was found that the additional elements in the 200+ token prompt (camera model, F-stop, ISO, Key Light angle, etc.) were reflected in the output.
Why Don’t Camera Specs Work?
This is speculative, but shooting parameters like Aperture: f/4 and ISO: 100 may appear in CLIP’s training data as camera metadata without being strongly associated with visual features of images. As a result, these specifications consume tokens without contributing to output.
Token Efficiency
| Original | Optimized | |
|---|---|---|
| Estimated tokens | 200+ (3+ chunks) | ~40 (within 1 chunk) |
| Core element reproduction | 6/6 | 6/6 |
| Quality difference | - | None observed |
The fact that the original — far exceeding the 75-token boundary — and the optimized version within a single chunk produce equivalent results demonstrates that what you write matters more than how much you write.
Note: Writing a specific camera model name does not make the image look like it was shot with that camera. Those tokens are better spent on pose and lighting direction.
Summary
The following elements can be safely removed from prompts:
| Element | Tokens Saved |
|---|---|
| Camera model names | 5-6 |
| Lens focal length & aperture | 3-4 |
| Shutter speed, ISO, white balance | 5-6 |
sharp focus, ultra-detailed skin texture | 5 |
| Redundant expressions of the same concept | Variable (10-20) |
| Individual Key/Fill/Rim Light details | 15-20 |
Reallocate those freed tokens to scene descriptions, poses, and lighting direction — elements that actually affect the image. That’s what prompt optimization is really about.
![[Verified] Image Generation Prompt Best Practices](/tips/prompt-best-practices/cover_0_0000_4517457392071889496.webp)
