Published: 2026-05-21
nanobanana prompts, reverse prompt engineering, AI image generation, Nano Banana Pro, prompt breakdown
Ever scroll Xiaohongshu or design sites and spot an amazing AI image—perfect lighting, killer typography—while the author never shares the prompt? Writing your own feels like a blind box: tweak until you’re exhausted and still miss the vibe.
Here’s a power move: let AI deconstruct the image and output Chinese prompts you can paste to recreate it.
Tool: Nana Banana Pro (Gemini 3 Pro image model)
Method: Structured breakdown template + image upload → precise recreation
What you need
- Nana Banana Pro chat (image upload supported)
- One reference image you want to recreate (text-heavy images hit 90%+ match)
- The breakdown template below (copy in full)
Breakdown template (copy in full)
Send this template to Nana Banana Pro, then upload your target image. Priorities P0 → P1 → P2 lock text first, then details, then lighting.
Note: The template below is in Chinese—copy it unchanged for best results with Nano Banana Pro.
你是专业的图片提示词拆解专家 任务：分析图片，输出可用于重新生成的中文提示词 拆解优先级： 【P0】文字内容 - 明确写出图中的具体汉字 - 描述字体风格（毛笔/霓虹/金属/3D等） - 说明排版布局 【P1】画面细节 - 补充模糊部分的合理推测 - 描述材质纹理（丝绸/金属/玻璃等） - 说明元素互动关系 【P2】光影渲染 - 光源类型和方向 - 渲染风格（写实/赛博朋克/复古等） 输出要求： • 必须用中文 • 文字描述放最前面 • 细节要具体可执行 • 不要分析过程，直接给提示词
| Priority | Dimension | What to capture |
|---|---|---|
| P0 | Text | Exact characters, font style, layout |
| P1 | Scene details | Materials, element relationships, reasonable fill-ins |
| P2 | Light & render | Light direction, photoreal / cyberpunk / retro, etc. |
Workflow (reusable)
Step 1: Send the template
Copy the full breakdown template and send it to Nana Banana Pro.
Step 2: Upload the image
Upload the image you want to reverse-engineer (posters, logos, and headline graphics work best).
Step 3: Get the prompt
The model returns a structured Chinese prompt ready for regeneration—no chain-of-thought, just the final prompt.
Step 4: Regenerate
Paste the prompt back into Nana Banana Pro (or any Chinese-capable image model) for a close recreation.
3 tips for higher fidelity
1. Clearer source text = tighter copy
The template’s P0 priority forces exact Chinese characters—huge for posters, logos, and titles.
2. More detail = closer match
If the first pass misses, manually tune materials, lighting, or layout and run again.
3. Iterate in layers, not one lucky roll
Lock text and composition first, then add P1 details and P2 lighting—more stable than vague adjectives.
Final thoughts
The hard part of AI images isn’t “can it draw”—it’s translating your taste into language the model understands. Master upload → structured breakdown → Chinese prompt recreation, and you move from blind boxes to controlled copies.