ningzimu/image-to-editable-ppt-skill

What it solves

This project provides a way to convert static images, PDFs, or image-based PowerPoint presentations into fully editable .pptx files. It solves the problem of having slides that are essentially just pictures, making it difficult to update text, move elements, or modify the layout without starting from scratch.

How it works

The tool normalizes input files into page-by-page tasks. It then uses a combination of AI agents (main agent and page workers) and tools to reconstruct the slides:

• Text Recovery: Uses OCR (specifically PaddleOCR-VL) to accurately determine text content, font size, and coordinates, restoring them as native PowerPoint text boxes.
• Shape Reconstruction: Simple geometric shapes are converted back into native PowerPoint shapes.
• Visual Assets: Complex visual elements are extracted as independent image assets, often using image generation and editing tools (like Codex's built-in image_gen or OpenAI-compatible APIs) to isolate foreground elements.
• Parallel Processing: For multi-page documents, the main agent distributes tasks to sub-agents (page workers) to process pages in parallel.
• Final Assembly: The tool collects the manifests from each page and assembles them into a final, ordered .pptx file, while preserving original slide notes.

Who it’s for

Users who need to transform non-editable slide content (like screenshots or PDF exports) into editable PowerPoint presentations for further refinement, or designers who want to replicate a visual design while keeping text editable.

Highlights

• Multi-format Support: Handles single images, multiple images, PDFs, and image-based PPTX files.
• Coded Reconstruction: Restores native PowerPoint tables, curves, dashed lines, and arrows rather than just placing images.
• Measurement-Driven: Uses OCR-driven measurements for precise text placement and sizing.
• Hybrid Strategy: Combines editable text, native shapes, and independent image assets for complex pages.
• Parallel Workflow: Supports concurrent page processing via sub-agents to speed up up the conversion of large decks.