image2editable turns images, PDFs, and screenshot-based PowerPoint slides into PowerPoint files that can be edited again. It is designed for courseware screenshots, design mockups, report pages, and image-based slides, reducing the work of recreating a page from scratch.

After conversion, you can edit recovered text, move separated visual elements, and keep refining the page in PowerPoint. When processing a mixed PPTX, existing editable content is retained in the file.

Conversion examples

Source	Editable result

For the best visual result in a 16:9 PowerPoint deck, use a 16:9 input image when converting a single image.

Features

Before you start

• This is a tool for rebuilding existing pages into editable PowerPoint files. It does not create a new presentation from an article or outline.
• ⚠️ Complex visuals are usually kept as movable image components. Their internal elements are not guaranteed to become native PowerPoint shapes.
• 🔒 Host Agent mode may send diagnostic images to the current host service. Before processing sensitive files, verify that the host service's data policy meets your requirements.

Quick start

Install with the skills CLI

npx skills add DSY-Xueai/image2editable --skill image-to-ppt

Let an Agent install it

Install the <image-to-ppt> skill from https://github.com/DSY-Xueai/image2editable.

After installation, describe the task to an Agent that supports vision, file access, and tool calls. Images, PDFs, and .pptx files can be pasted or attached in the chat, or provided as local paths:

# Codex
$image-to-ppt Convert input.pptx to an editable PPTX and preserve native objects that are not selected for reconstruction.
$image-to-ppt Convert input.png to an editable PPTX.
$image-to-ppt Convert <input.pdf> to an editable PPTX.

# Claude Code
/image-to-ppt Convert input.pptx to an editable PPTX and preserve native objects that are not selected for reconstruction.
/image-to-ppt Convert input.png to an editable PPTX.
/image-to-ppt Convert <input.pdf> to an editable PPTX.

With a full repository or an installed image2editable Runtime, the Skill automatically prepares the pinned dependencies, OCR, and runtime models, then completes the conversion through the current Host Agent. A standalone Skill installation requires absolute local paths in SAM2_MODEL, LAMA_MODEL, and GROUNDING_DINO_MODEL; if any path is missing, the Skill lists the missing variables and stops.

Project layout

image2editable/
├── .claude-plugin/            # Claude Code plugin manifest
│   └── plugin.json
├── .github/                   # CI, issue forms, and PR template
├── docs/
│   └── images/                # README image assets
├── image2editable/            # Unified CLI, runtime, and conversion modules
├── scripts/                   # Recognition, reconstruction, and PPTX/PSD assembly modules
├── skills/
│   ├── image-to-ppt/          # Installable image-to-PPT Skill
│   └── image-to-psd/          # Compatible image-to-PSD Skill
├── tests/                     # Automated tests
├── third_party/
│   └── licenses/              # Third-party license materials
├── .gitignore
├── CITATION.cff               # Citation information
├── image_to_ppt.py            # Legacy image-only pipeline; not the recommended entry point
├── image_to_psd.py            # Compatible image-to-PSD entry point
├── LICENSE                    # MIT license
├── pyproject.toml             # Python package and CLI configuration
├── README.md                  # Chinese documentation
├── README_EN.md               # English documentation
├── requirements.txt           # Core dependencies
└── THIRD_PARTY_NOTICES.md     # Third-party dependency and license notices

Known limitations

• ⚠️ Review complex pages manually. Decorative text, dense tables, gradients, and complex illustrations may not be restored pixel for pixel. Check text, component positions, and layout before delivery.
• Clear text and regular backgrounds generally reconstruct more reliably. Decorative text, dense tables, gradients, and complex illustrations are not guaranteed to match pixel for pixel.
• 💳 Host Agent consumes the current model's token/context allowance. Complex pages may require several diagnostic and repair rounds; actual usage depends on the Agent, model, and page complexity.
• ⏱️ Multi-page PDFs, complex pages, and high-resolution images take longer. Each page goes through OCR, visual separation, reconstruction, and quality checks, with up to five repair rounds. Host mode also waits for Agent visual decisions.

Supported inputs

See THIRD_PARTY_NOTICES.md for third-party dependencies and licenses, and CITATION.cff for citation information.

License

MIT