px-image2pptx

name: px-image2pptx description: > Convert static images (slides, posters, infographics) to editable PowerPoint files. Pipeline: OCR detects text → classical CV textmask detects ink pixels → mask-clip ANDs with OCR bboxes (preserves illustrations) → LAMA inpaints clean background → python-pptx assembles editable text boxes with auto-scaled fonts and detected colors. Trigger on 'convert image to pptx', 'make slide editable', 'image to powerpoint', 'extract text from slide as editable', 'reconstruct slide', or when the user has a slide/poster image and wants an editable .pptx file. metadata: author: Jade Liu source: https://github.com/JadeLiu-tech/px-image2pptx risk: low

px-image2pptx: Image to Editable PowerPoint

What It Does

Converts a static image into an editable .pptx file where every text element is a selectable, editable text box over a clean inpainted background. The full pipeline runs in 8-16 seconds:

1. OCR (PaddleOCR) — detects text regions with bounding boxes and content
2. Textmask (classical CV) — finds text ink pixels via adaptive thresholding
3. Mask-clip — ANDs textmask with OCR bboxes to preserve non-text elements
4. Inpaint (LAMA) — reconstructs masked regions with neural inpainting
5. Assemble — places editable text boxes with auto-scaled fonts and detected colors

When to Use This

Installation

git clone https://github.com/JadeLiu-tech/px-image2pptx.git
cd px-image2pptx
pip install -e ".[all]"

Usage

CLI

# Full pipeline
px-image2pptx <image> -o <output.pptx>

# Chinese text
px-image2pptx <image> -o <output.pptx> --lang ch

# Skip inpainting (solid-bg slides)
px-image2pptx <image> -o <output.pptx> --skip-inpaint

# Pre-computed OCR (skips PaddleOCR dependency)
px-image2pptx <image> -o <output.pptx> --ocr-json text_regions.json

# Keep intermediate files
px-image2pptx <image> -o <output.pptx> --work-dir ./debug/

Python API

# One-line conversion
from px_image2pptx import image_to_pptx
report = image_to_pptx("slide.png", "output.pptx")

# With options
report = image_to_pptx(
    "slide.png", "output.pptx",
    lang="ch",            # Chinese OCR
    sensitivity=16,       # textmask sensitivity
    dilation=12,          # textmask dilation
    skip_inpaint=False,   # True for solid-bg slides
    work_dir="./debug/",  # save intermediates
)

# Step-by-step
from px_image2pptx.ocr import run_ocr
from px_image2pptx.textmask import compute_masks
from px_image2pptx.inpaint import inpaint
from px_image2pptx.assemble import assemble_pptx

regions = run_ocr("slide.png", lang="ch")
tight, clipped, dilated = compute_masks(cv2.imread("slide.png"), regions)
background = inpaint(cv2.cvtColor(cv2.imread("slide.png"), cv2.COLOR_BGR2RGB), dilated)
report = assemble_pptx("slide.png", regions, "output.pptx", "background.png", tight)

CLI Options

OCR JSON Format

{
  "text_regions": [
    {
      "id": 0,
      "text": "Hello World",
      "confidence": 0.95,
      "bbox": {"x1": 100, "y1": 50, "x2": 400, "y2": 90}
    }
  ]
}

Key Design Decisions

Font auto-scaling (90-94% of bbox width)

Uses PIL ImageFont for accurate Latin text measurement, 1.0x em for CJK. Bidirectional: shrinks if text overflows, grows if text is too small.

Text color from tight mask

Samples ink pixel colors from the pre-dilation mask (actual ink, no background bleed). Falls back to contrast-based sampling for light-on-dark text where textmask misses.

Column-aware line grouping

Two-pass: vertical proximity, then horizontal gap split (3x median height). Prevents merging left/right column text into one wide text box.

OCR-guided masking

Raw textmask detects ANY dark pixels (including illustration outlines, borders). Mask-clip ANDs with OCR bboxes so only confirmed text gets inpainted. Removes 30-90% of false positive mask pixels depending on image content.

Dependencies

Models

Models are downloaded automatically on first use (~370 MB total).

Performance

Limitations — When to Warn the User

Before running the pipeline, assess the input image and set expectations:

When NOT to use this skill

• The user wants to extract individual assets (icons, logos, illustrations) → use px-asset-extract
• The user wants to read text from an image without creating a PPTX → use OCR directly
• The input is already a .pptx file → use the pptx skill to edit it directly
• The image has no text (pure photo, illustration) → there's nothing for this pipeline to do