Visual Asset Cards from PDF

Purpose

Convert architectural project visualization references (PDF brochures, image folders, WeChat attachments) into a reusable visual standard for the rendering team. The output is three asset cards:

1. Main Anchor Card — what the project looks like (visual DNA)
2. Style Guard Card — must-haves and taboos so outputs do not drift
3. Scenario Expansion Card — reusable scene-specific prompts for batch production

When to Use

Use this skill when the user:
- References a PDF or image set of project renderings/photos and asks to "extract style", "generate asset cards", "make rendering standards", or "help the team render consistently".
- Wants to turn a competitor or benchmark project into an internal visual prompt library.

Required Inputs

• Path to PDF(s) or image folder(s)
• Project name / location (extract from OCR if needed)
• Any known constraints (e.g., no red/gold, low-sat, hotel-like, etc.)

Workflow

Step 1: Ingest & Convert

1. If input is PDF:
   - Check page count, size, and whether it is a scanned/image-only PDF (no text layer).
   - Convert all pages to PNG at a workable width (around 1600 px) using pymupdf (fitz).
   - Generate thumbnail JPEGs (around 400 px) for quick browsing.
   - Run OCR (rapidocr_onnxruntime) to extract any text; save to _ocr.txt.
2. If input is an image folder:
   - Create thumbnails if absent.
3. Create a browsable HTML index (_index.html) pairing each page thumbnail with its OCR text.

Step 2: Visual Research

1. Open the thumbnail index and browse every page. Do not skip pages because OCR is empty — these are often pure rendering pages.
2. Categorize pages by scene type:
   - Main entrance / gate / porte-cochère
   - Clubhouse / lobby / canopy
   - Covered walkway / colonnade
   - Courtyard garden / water feature
   - Landscape terrace / outdoor furniture
   - Building facade detail
   - Signage wall / branding moment
   - Interior-to-exterior framed view
   - Driveway / approach
3. Record repeating patterns:
   - Time of day (e.g., blue hour dusk)
   - Weather / lighting (soft overcast, warm interior glow)
   - Material palette (stone, metal, glass, water, plants)
   - Color palette (warm grey, champagne gold, dark granite, muted green)
   - Camera angle (eye-level 1.2–1.6 m, 24–35 mm, verticals corrected)
   - Composition rules (symmetry, framing, foreground occlusion, leading lines)

Step 3: Distill Visual DNA

Synthesize findings into a one-paragraph style definition plus a table of dimensions:

Step 4: Generate the Three Asset Cards

Write the output to a Markdown file named _asset_cards_<project_slug>.md and an HTML preview named _asset_cards_<project_slug>.html.

Card 1: Main Anchor

• One-sentence style definition.
• Feature table with the seven dimensions above.
• Bilingual prompt (English and Chinese) that can be dropped into Midjourney / Stable Diffusion / D5 / Enscape as a master prompt.

Card 2: Style Guard

• Must-haves (time, color, material combo, lighting, plants, camera, mood).
• Taboos (e.g., no plastic plants, no dead-black shadows, no oversaturated colors).
• Pre-delivery self-check list (6–8 questions).

Card 3: Scenario Expansion

• Provide 6–10 scene-specific prompts covering the categories observed.
• Each prompt includes: scene name, composition notes, English prompt, Chinese prompt.

Step 5: Deliver

1. Present the Markdown and HTML files using present_files.
2. Tell the user the files are ready and offer follow-up: scene-specific extraction, visual comparison with another project, or rendering-parameter cheat sheet.

Output Conventions

• Use project slug in filenames: _asset_cards_<project_slug>.md and .html.
• Keep prompts concise but specific enough to reproduce the look without the reference image.
• Include both English and Chinese prompts because rendering tools often prefer English, while the team discussion is in Chinese.
• The HTML preview should use a clean, white/light theme with readable code blocks and tables.

Bundled Resources

• None required; all operations use standard Python libraries (fitz, rapidocr_onnxruntime) and built-in tools.