ppt-image-to-editable-ppt

name	ppt-image-to-editable-ppt
description	Convert PPT slide screenshots or exported slide images into editable PowerPoint decks. Use when Codex needs to extract image/icon/material assets from one or many slide images as separate PNGs, rebuild the slides with editable text boxes, native shapes, and movable picture objects, batch-process multiple page images, and merge the reconstructed pages into a complete .pptx file.

PPT Image To Editable PPT

Overview

Turn slide images into a reconstructed, editable PowerPoint deck. Text stays as real editable PowerPoint text. Icons, logos, photos and decorative bands come out as separate movable PNG picture objects. Cards, frames, circles and rules are rebuilt as native shapes/lines where the source geometry is simple enough.

This skill is for slide screenshots or exported slide images. If a native .pptx already exists, edit that directly.

Supported input formats use page_NN.<ext> naming and include PNG, JPG/JPEG, WebP, BMP, and TIF/TIFF. Extracted assets and previews are written as PNG.

Dependencies

Run the bootstrap script from the repository root:

bash scripts/bootstrap.sh

It installs the Python packages in requirements.txt, local OCR tools, LibreOffice, Poppler, and model caches where the platform supports automatic installation. On managed systems, install these manually and use bash scripts/bootstrap.sh --no-system.

Run Directory

Every job goes under:

output/<name>_<YYYYMMDD>/
├── slides.pptx
├── qa.json
├── previews/
├── ocr/
├── inventory/
├── manifests/
├── layouts/
├── assets/
└── debug/

Workflow

Prepare source images:

slides/
├── page_01.png
├── page_02.jpg
└── page_03.webp

Run the one-command pipeline:

RUN="output/demo_$(date +%Y%m%d)"
SRC="slides"

python scripts/convert.py --source "$SRC" --work-dir "$RUN"

For debugging, the same flow can be split into the three lower-level steps:

python scripts/ocr/prepare_ocr.py \
  --source-dir "$SRC" \
  --work-dir "$RUN"

python scripts/ocr/ocr_review_apply.py --work-dir "$RUN"

python scripts/build_deck.py \
  --source-dir "$SRC" \
  --work-dir "$RUN"

convert.py creates or uses a run directory, normalizes single-image or non-page_NN sources when needed, then calls the three lower-level steps. prepare_ocr.py loads PaddleOCR once, runs OCR across all pages, builds review packets for uncertain entries, and pre-fills corrected_text with local consensus picks. ocr_review_apply.py merges those picks into the OCR JSON files. build_deck.py runs erase, inventory extraction, layout generation, deck assembly, preview-based font-size calibration, preview-based text position calibration, QA inspection, and preview rendering.

Optional Review

After build_deck.py, check whether it reports OCR entries that may benefit from manual review. If the user wants the extra pass:

1. Open ocr/page_NN.ocr_review.annotated.png.
2. Edit ocr/page_NN.ocr_review.json.
3. Set corrected_text to the desired text, or to "" when the OCR detection is not real editable text.
4. Rerun:

python scripts/ocr/ocr_review_apply.py --work-dir "$RUN"
python scripts/build_deck.py --source-dir "$SRC" --work-dir "$RUN"

Useful Flags

• --pages 1,4,8: process only selected pages.
• --skip-render: skip LibreOffice preview generation and default text calibration.
• --skip-calibration: skip preview-based text size/position calibration.
• --font-calibration-iterations 1: set font-size calibration passes.
• --calibration-iterations 4: set text position calibration passes.
• --detect-tables: enable optional native table reconstruction.
• --icon-review: emit icon-vs-text review packets.
• --icon-decisions: apply filled icon review decisions on a rerun.

These flags can be passed to scripts/convert.py.

Validation Standard

A job is done when:

• meaningful visual assets are present as PNG picture objects or native shapes;
• ordinary text is editable in the PPTX;
• extracted PNG assets are independent objects, not a flattened full-page background;
• inspect_pptx.py reports no failures;
• previews have been compared with the source images, or the final response states why preview rendering was skipped or unavailable.

Final responses should include the final .pptx path, the preview directory, the QA report, and any known fidelity differences.