guopenglin/image2ppt

Chinese | English

DeckWeaver can reconstruct images output from GPT, Gemini, and others into editable PowerPoint files. It can be used as an agent skill or directly as a command-line tool without large model intervention.

Almost all text on images will be restored into editable PPT text boxes, and icons, logos, images, and decorative elements will be broken down into independent image objects.

Advantages

• Almost all text and icons can be edited, moved, or replaced in PowerPoint.
• Very low token consumption (command-line tools can even consume no tokens): The main workflow uses local OCR and image algorithms, so there’s no need to repeatedly hand over entire pages to cloud multimodal models.
• Fast generation speed: batch OCR reuses hot models, with subsequent page pipelines automatically completed.
• No need for additional cloud APIs: OCR, image segmentation, PPTX generation, and preview checking all run locally.
• Supports common bitmap inputs: PNG, JPG/JPEG, WebP, BMP, TIF/TIFF.
• Supports PDF input: directly read PDF text layers, skip OCR, and restore text to 100% (see below).

Example of the effect

The left side of the image below shows the original input (a privacy calculation schematic generated by GPT), and the right side shows a preview of the PNG rendered back to PNG by LibreOffice after DeckWeaver’s automatic reconstruction. All text is an editable text box, and the MPC center lock, connecting cables, etc., have been split into independent, movable PNG objects.

Enter the original image Rebuilding PPT (preview)

Get Started Quickly

Method 1: Use as a Skill / Agent tool

1. Clone the project, or directly clo it to your skill directory:

2. Open the project directory using Codex, Claude Code, or another local agent.

3. Tell the agent the image or image folder you want to convert, for example:

You can also specify a single image:

Note When running for the first time, the agent will first execute bash scripts/bootstrap.sh to install dependencies, which may take some time. The final result is at output/ /slides.pptx.

Method 2: Use as a standalone command-line tool

Suitable for the following types of users:

• Don’t want to spend tokens

• Minimal requirements for text accuracy (since large models do not intervene in text recognition)

• Want to use the command line to achieve large-scale recognition

bootstrap.sh will install Python dependencies, local OCR dependencies, LibreOffice/Poppler preview tools, and pre-download model caches. macOS and common Linux distributions can be used directly; Windows or managed environments can refer to requirements.txt for manual installation of dependencies.

Then run with one click:

You can also directly process individual images:

Or you can send a PDF directly, with each PDF page corresponding to a page of the PPT: