VoxelCubes/PanelCleaner

Panel Cleaner – AI‑powered comic‑panel text removal

What it does – Panel Cleaner is a Python‑based application that automatically detects text inside speech‑bubbles of manga/comic pages, creates precise masks, and either removes the text (leaving a clean bubble) or extracts the text via OCR. It uses a pre‑trained text‑detection model, LaMa for in‑painting, and optional MangaOCR/Tesseract for OCR. The tool can run head‑less from the command line or through a Qt GUI, supports CUDA acceleration, and works on Windows, macOS, and Linux.

Key features

• Detects Japanese and English text, generates high‑precision masks, and expands them to cover nearby artifacts.
• In‑paints bubbles that cannot be cleanly masked using the LaMa model.
• Saves either the cleaned image, a transparent mask layer, or both.
• Batch processing of folders and full‑directory trees.
• Fully configurable via plain‑text profiles (INI‑style) – every parameter (mask growth, thresholds, etc.) can be tweaked and saved.
• Built‑in OCR: MangaOCR for Japanese, Tesseract for English/Japanese, with an interactive editor in the GUI.
• No internet required after the initial ~500 MiB model download.
• Optional CUDA support for faster processing (only when installed via pip, not in the pre‑built binaries).
• Multilingual UI (English, German, Bulgarian, Spanish) and community translation support.

How to get it

1. Pre‑built binaries – download the latest release from the GitHub Releases page (recommended for most users).
2. Python package – pip install pcleaner (installs both CLI and GUI) or pip install pcleaner-cli for command‑line only.
3. AUR – yay -S panelcleaner (installs in a pipx environment, giving easy CUDA handling).
4. Flatpak – io.github.voxelcubes.panelcleaner from Flathub (no CUDA).
5. Docker – build the provided Dockerfile and run the container, mounting a host folder with the images.

Typical command‑line workflow

# Clean a few files, saving cleaned images
pcleaner clean page1.png page2.png

# Clean a folder, keep only the mask layers
pcleaner clean -m my_folder

# Use a custom profile
pcleaner clean -p my_profile folder/

# Run OCR and write results to a text file
pcleaner ocr my_folder --output-path=out.txt

The GUI can be launched with pcleaner gui or pcleaner-gui and offers visual previews of masks, analytics, and an OCR editor.

Limitations

• Only Japanese and English are supported for cleaning; OCR works for Japanese (via MangaOCR) and English (via Tesseract).
• Masks are grayscale only, so they work on white/black/gray bubbles but not on colored bubbles.
• The AI detector is imperfect – 2‑8 % of bubbles may be missed or incorrectly masked, and the tool will skip bubbles it cannot clean confidently.
• Pre‑built binaries lack CUDA; GPU acceleration requires the pip installation with a compatible PyTorch build.

License – GNU GPL‑v3 (free to use, modify, and redistribute under the same license).

Who might benefit – translators, scan‑lators, or anyone who needs to strip text from manga panels quickly before re‑lettering, as well as researchers needing clean bubble images for downstream computer‑vision tasks.