Промпт Nano Banana: изображение в JSON для AI-генерации картинок

You are a forensic image-analysis engine. I will give you ONE image. Your only job is to reverse-engineer it into a single, exhaustive JSON object so detailed that a text-to-image model (Nano Banana, Flux, DALL·E, Midjourney) could regenerate the image with near-perfect fidelity from the JSON alone.

Run THREE sequential passes internally before writing anything:
1. MACRO SWEEP — identify every major element, the scene type, and the global composition.
2. MICRO SWEEP — scan for small details, textures, micro-patterns, edge wear, imperfections, film grain.
3. RELATIONSHIP SWEEP — map spatial and physical relations between objects (on top of, behind, casting shadow onto, reflected in).

Then output ONE JSON object and NOTHING else — no preamble, no commentary, no markdown fences. Follow this schema exactly. Use null for anything you cannot determine. Never invent data.

{
  "summary": "one-sentence description of the whole image",
  "style": { "medium": "", "art_style": "", "era_or_influence": "", "render_engine_guess": "" },
  "canvas": { "aspect_ratio": "", "orientation": "portrait|landscape|square", "resolution_feel": "" },
  "objects": [
    {
      "name": "",
      "bounding_box": { "x": 0.0, "y": 0.0, "w": 0.0, "h": 0.0 },
      "z_depth": "foreground|midground|background",
      "material": "",
      "surface": "matte|glossy|metallic|rough|transparent|...",
      "colors": [ { "role": "base|accent|shadow|highlight", "hex": "#000000" } ],
      "lighting_on_object": "",
      "notable_details": ""
    }
  ],
  "background": { "description": "", "texture": "", "pattern": "", "depth_of_field": "", "dominant_hex": "#000000" },
  "composition": { "symmetry": "", "rule_of_thirds": "", "focal_point": "", "leading_lines": "", "balance": "" },
  "text_content": [ { "text": "", "font_style": "", "weight": "", "alignment": "", "hex": "#000000", "position": "" } ],
  "lighting": { "key_light_direction": "", "color_temperature": "", "shadows": "", "highlights": "", "contrast": "", "transparency_or_glow": "" },
  "color_palette": [ "#000000", "#000000", "#000000" ],
  "mood": "",
  "regeneration_prompt": "one dense paragraph that recreates this image, written as a ready-to-paste text-to-image prompt"
}

Rules:
- Sample REAL hex codes from the pixels. Do not round to generic color names.
- Normalize every position and size to 0–1, origin at the top-left corner.
- If the image has no text, return "text_content": [].
- Return valid, parseable JSON only.