Comic Ocr

comic-ocr

OCR comic strips on macOS using Apple's Vision framework, then clean up the transcripts with Claude.

Two scripts, meant to be run in sequence:

1. **`ocr.swift`** — Extracts text from a directory of comic strip images using Apple Vision's highest-accuracy recognition mode. Outputs a JSON file with one entry per image.

2. **`clean.py`** — Sends the raw OCR through Claude (via `claude -p`) to fix the usual OCR mess: garbled characters, ALL-CAPS normalization, metadata stripping, dialogue structure cleanup. Processes in parallel batches and writes results incrementally.

Requirements

• macOS (uses Apple's Vision framework — no cloud OCR, everything runs locally)
• Swift (comes with Xcode or Xcode Command Line Tools)
• Python 3
• Claude Code CLI (claude on your PATH)

Usage

Step 1: OCR

# Compile (one-time)
swiftc ocr.swift -o ocr -framework Vision -framework AppKit

# Run
./ocr /path/to/strips transcripts.json

Progress is logged to `ocr_progress.log` in the output directory. You can `tail -f` it.

The output JSON looks like:

[
  {
    "filename": "0.gif",
    "text": "SOMETHING POSITIVE R*K NICHOLLAND\nKNIFE GOES\nIN...\nOKAY, DAVAN, WE\nOUT FOR METE..."
  },
  {
    "filename": "1.gif",
    "text": "..."
  }
]

Raw OCR is noisy — comic strips typically come through as ALL-CAPS with the title, copyright, and URL baked into every image. The text may have broken words and garbled characters.

Step 2: Clean up with Claude

python3 clean.py transcripts.json transcripts_clean.json

Or just `python3 clean.py` if your files are named `transcripts.json` (output defaults to `transcripts_clean.json`).

This sends batches of 50 strips at a time to Claude Haiku, 4 batches in parallel. Progress is logged to `transcripts_clean.log`. Results are written to the output file incrementally as each batch completes.

The cleanup fixes:

• Strips comic title headers, author bylines