Gemini + NanoBanana: Prompt Reverse Engineering for Image-to-Image
Extract structured style JSON with Gemini, then reshape uploads with NanoBanana Pro image-to-image—more accurate than blind style transfer; includes templates and official nanobanana入口.
When you love a reference image but cannot describe its “look”, structure the style first, then let NanoBanana Pro do image-to-image. This usually beats “only feeding the reference” because you control color, composition, lighting, and texture explicitly.
Step 1: Reverse to JSON (style contract)
Ask Gemini to output JSON only after you attach the reference:
Extract visual style from the image as strict JSON only (no extra text). Include fields such as colors, composition, visual_effects, overall_style. Use empty arrays/strings when unknown.
Step 2: Reshape your image with NanoBanana Pro
Upload your target image and run:
You are an image director. Reshape the uploaded image to match the aesthetic system described by the JSON below. Keep subject structure unless instructed otherwise. JSON: {…}
Why JSON helps
- Blind style transfer often copies colors but misses lighting logic
- JSON lets you edit one field at a time (e.g., lighting only)