Conclusion
- NB2’s Japanese natural language must be decomposed into English tag sequences for z-image-turbo — comma-separated keywords are more effective than sentences
- Composition elements must be explicitly specified or they won’t work —
from below,shot through window, etc. need to be included in the prompt - Always place color + subject adjacent — write them as a set like
dark beige brown medium hairto prevent attribute leakage - “Lowering quality” instructions work in z-image-turbo too — retro/amateur feel specifiers like
low quality photo lookandfilm grainare surprisingly effective - Backlighting is a strength of z-image-turbo — the combination of
backlighting+rim light+silhouetteconsistently produces beautiful backlit images
Popular Nano Banana 2 (Gemini) prompts were converted for z-image-turbo and generated. Even prompts with the same intent require different writing styles because the model mechanisms differ.
This article publishes the conversion process and actual generation results for 10 themes.
Differences Between Nano Banana 2 and z-image-turbo
Understanding why conversion is necessary.
| Item | Nano Banana 2 | z-image-turbo |
|---|---|---|
| Base technology | Gemini (LLM) | Stable Diffusion-based (CLIP + U-Net) |
| Prompt language | Japanese OK (natural language understanding) | English recommended (tag sequences are effective) |
| Context understanding | High (understands sentence meaning) | Low (token-level matching) |
| Token limit | Long text OK | 77 tokens (~50-60 English words) |
| Weight of early content | Uniform | Earlier positions have stronger influence |
| Color attribute leakage | Less common | Prone to occurring (red, dress, blue, ocean → colors mix) |
See prompt basics and how CLIP works for more detail.
Conversion Rules
- Japanese → decompose into English tag sequence (comma-separated, not sentences)
- Lead with Subject (
1girl, 20yo japanese woman) - Negations → separate into negative prompt (※ In z-image-turbo, CFG=1.0 means negative prompts don’t work)
- Emphasize important composition/lighting (
(from below:1.4)etc.) - Place color + subject adjacent (prevents attribute leakage:
red dress○,red, dress×) - Remove camera model names (little effect in z-image-turbo, confirmed by experiment)
1. Classic Gravure Standing Pose
A healthy gravure photo at poolside.
Conversion notes: Micro-adjustments like “showing a long neck” and “fingers carefully arranged” have weak effect in CLIP, so they were omitted. magazine gravure style unifies the gravure feel.
2. Hair and Makeup Focus Portrait
Hair color and makeup color specifications are the main focus. Attribute leakage prevention (colors mixing) is critical.
Conversion notes: Placing color + subject adjacent like dark beige brown medium hair to prevent attribute leakage. Makeup colors also written as sets like beige gold eyeshadow.
3. Emotional Composition (Library × Rib Knit)
Creating atmosphere through a “world-set” of place × outfit × light.
Conversion notes: The “world-set of place × outfit” technique is explained in the composition 4-axis theory. The cream knit + check skirt + wooden chair + bookshelf texture contrast (soft material × hard wood) enhances the beauty. looking away is placed in the positive prompt (not negative) to direct the gaze away.
4. Amateur SNS Style (Intentionally Lowering Quality)
A reverse approach where “not being too polished” creates realism.
Conversion notes: While normally high quality, professional, etc. are added, this theme emphasizes low quality photo look instead. professional lighting, studio background, perfect composition would also go in the negative prompt to eliminate the professional feel.
5. 1990s Retro Digicam Style
Using flash + grain + overexposure to create a retro feel.
Conversion notes: Emphasizing (on-camera flash:1.4) and (film grain:1.3) to strengthen the core retro elements. disposable camera aesthetic unifies the overall tone. The red lantern light from the side and flash from the front creates a dual light source effect.
6. Low-Angle (from below) for Impact
Low angle to express confidence and power.
Conversion notes: Without emphasizing (from below:1.4), the angle may not register. 24mm wide angle exaggerates wide-angle perspective distortion, and looking down at viewer creates a looking-down gaze.
7. High Angle (from above) for Cuteness
Same setting as #6 but only the angle changed. Comparing the impression difference.
Conversion notes: Same subject and outfit as #6, but changing from below → from above, standing → sitting, confident → gentle lonely is enough to flip the impression 180 degrees. Among the composition 4 axes, changing only “camera height” can switch between powerful and cute.
8. Backlighting + Transparency
Expressing silhouette and rim light with backlighting.
Conversion notes: Triple specification of (strong backlighting:1.4) + golden rim light + silhouette face forces the backlit look. water surface reflections, sparkles adds water surface light. ethereal atmosphere unifies the overall transparency. In z-image-turbo, backlighting tends to clearly separate the light beams from the subject’s silhouette.
9. Office Fluorescent Light Realism
Office setting with fluorescent + window light combination.
Conversion notes: The NB2 version specified “Canon EOS R5” — deleted. Only the lens spec 50mm f/1.4 was kept. (fluorescent light:1.2) creates the office fluorescent feel, while afternoon window light adds natural window light mix.
10. Foreground Blur + Through-Window Three-Layer Composition
A cafe shot through a window in three layers (foreground blur + subject + background blur).
Conversion notes: Double emphasis with (shot through window glass:1.3) and (blurry foreground:1.3) to force the through-window shooting. z-image-turbo beautifully reproduced the three-layer composition (foreground blur + subject + background blur). The glasses specification was also accurately reflected.
Overall Assessment of Conversions
Well Reproduced
| Theme | Reproduction | Notable Strengths |
|---|---|---|
| #3 Library | ★★★★★ | World-set of place × outfit × light was perfect |
| #5 Retro Izakaya | ★★★★★ | Flash + grain + red lantern atmosphere |
| #8 Backlit Beach | ★★★★★ | Silhouette + water surface reflection + transparency |
| #10 Window Cafe | ★★★★★ | Three-layer composition reproduction was impressive |
| #1 Poolside | ★★★★☆ | Swimsuit and pose accurate. Resort feel present |
What NB2→z-image-turbo Conversion Taught Us
- Composition elements must be explicitly specified:
from below,shot through windowand similar composition specifications won’t register if not included. Note that weight syntax like(from below:1.4)— the numeric strength variation — has not been confirmed - Color + subject must always be adjacent: Write as a set like
dark beige brown medium hair - “World-sets” work in SD-based models too: The place × outfit × light triangle also functions through CLIP token matching
- “Lowering quality” instructions work in SD-based models: Retro/amateur feel specifiers like
low quality photo lookandfilm grainare surprisingly effective - Backlighting is a strength of z-image-turbo: Beautiful backlighting consistently comes from the
backlighting+rim light+silhouettecombination
Related Articles
⚠ 関連記事が見つかりません: /en/tips/prompt-basics
⚠ 関連記事が見つかりません: /en/tips/prompt-design-thinking
⚠ 関連記事が見つかりません: /en/tips/prompt-iteration-bikini
⚠ 関連記事が見つかりません: /en/reviews/z-image-turbo-review
⚠ 関連記事が見つかりません: /en/papers/clip
![[Tested] Can You Generate Kuchisake-onna? Mask On/Off Pairs and 3 Slit-Mouth Phrasings on the Same Seed](/tips/kuchisake-onna-mask-expression-test/c01_lacer_s1.webp)
![[Tested] Can You Generate Yuki-onna? Pale Skin, Visible Cold Breath, and a Loosened White Juban](/tips/yuki-onna-pale-skin-test/b01_juban_s2.webp)