PanelCleaner: The Secret Tool Manga Translators Are Hiding

What if I told you that the most soul-crushing part of manga translation—hours of meticulously erasing Japanese text from speech bubbles—could be reduced to mere minutes? That the repetitive, wrist-destroying work that drives cleaners to quit could be handed off to an AI that actually understands what it's looking at?

If you've ever spent a Saturday night hunched over Photoshop, carefully cloning stamp tool over tiny pixels of text in a white bubble, only to realize you missed a speck and have to start over... this article is your escape hatch. The manga scanlation community has been bleeding talent for years because cleaning is brutal, monotonous, and thankless work. But what if the machines could handle the monotony while humans focus on the art?

Enter PanelCleaner—an open-source, AI-powered tool that automatically detects and removes text from manga panels with surgical precision. No more hand-drawing masks. No more guessing where the text ends and the artwork begins. This isn't some vaporware promise; it's a battle-tested tool already saving hundreds of hours for translators worldwide. And the best part? It's completely free, runs offline, and you can start using it in under ten minutes.

What Is PanelCleaner?

PanelCleaner is an AI-powered desktop application created by VoxelCubes that automates the most tedious step in manga localization: cleaning text from speech bubbles and panels. Built on top of the Comic Text Detector machine learning model, it doesn't just find text—it generates precise masks, expands them intelligently to catch artifacts, and can even inpaint bubbles that resist simple masking.

The project emerged from a clear pain point in the scanlation pipeline. Traditional cleaning requires human operators to manually identify text regions, create selections, and carefully erase content without damaging underlying artwork. For a typical manga chapter with 20-30 pages and hundreds of bubbles, this process consumes 2-4 hours of concentrated effort. PanelCleaner collapses this to minutes.

What makes PanelCleaner particularly noteworthy is its conservative, safety-first approach. Unlike crude "erase everything" tools that damage artwork, PanelCleaner is designed to skip bubbles it can't clean confidently rather than risk ruining the image. This philosophy—"do no harm"—makes it trustworthy for production workflows where quality matters more than perfect automation.

The tool has gained significant traction since its release, available through multiple distribution channels: pre-built binaries for Windows/Linux, PyPI packages for Python^{↗ Bright Coding Blog} users, Arch User Repository (AUR) for Arch Linux enthusiasts, Flatpak for containerized deployment, and even Docker^{↗ Bright Coding Blog} for cloud workflows. This flexibility reflects mature open-source project management and broad community adoption.

Key Features That Make PanelCleaner Insane

Let's break down what makes this tool genuinely powerful for production manga cleaning workflows:

AI-Powered Text Detection with Smart Masking The core engine uses deep learning to locate text bubbles and generate pixel-precise masks. But it doesn't stop there—the system expands masks (shown in purple in visualizations) to capture nearby undetected text and JPEG compression artifacts that would otherwise leave ugly remnants. For tight-fitting masks, it applies edge denoising (blue visualization) to clean borders without touching surrounding artwork.

LaMa Inpainting for Stubborn Bubbles Some bubbles can't be simply masked out—complex backgrounds, gradients, or detailed artwork behind text require intelligent reconstruction. PanelCleaner integrates Simple Lama Inpainting with a fine-tuned anime/manga model to fill these gaps convincingly.

Symbol and Number Detection Here's a subtle but brilliant feature: the AI ignores bubbles containing only symbols or numbers. Why? Because these don't need translation. This saves enormous time by eliminating false positives that plague generic text removal tools.

Full GUI with Theme Support Not everyone loves the command line. PanelCleaner offers a polished graphical interface with dark, light, and system theme options—essential for those late-night cleaning sessions.

Offline Operation Once you download the ~500MB model data on first launch, everything runs locally. No API keys, no subscription tiers, no internet dependency. This is crucial for handling copyrighted material and maintaining workflow reliability.

CUDA Acceleration For users with NVIDIA GPUs, installing via pip with proper PyTorch CUDA support unlocks dramatic speed improvements. The project's demonstration shows 46 images processed in real-time—try matching that by hand.

Batch Processing and Directory Handling Point it at a folder, walk away, return to cleaned images. The tool creates organized output directories automatically.

Profile System for Fine-Tuned Control Every parameter is exposed through editable configuration profiles. Adjust mask growth steps, expansion pixels, denoising intensity, and dozens more variables. Save presets for different manga styles or quality requirements.

Integrated OCR Pipeline Beyond cleaning, PanelCleaner runs Japanese OCR via MangaOCR and supports Tesseract for English/Japanese text extraction. Review and edit OCR output interactively in the GUI before exporting to translation files.

Real-World Use Cases Where PanelCleaner Dominates

1. High-Volume Scanlation Groups

Professional scanlation teams processing multiple series weekly face impossible deadlines. PanelCleaner's batch processing and CUDA acceleration allow a single operator to prep entire chapters while translators work on script. The conservative skip-behavior means quality checkers spend less time fixing AI mistakes.

2. Official Localization Workflows

Licensed manga publishers increasingly adopt AI-assisted tools to reduce costs and accelerate releases. PanelCleaner's offline operation addresses security concerns with pre-release materials, while its profile system enables standardization across production teams.

3. Academic and Archival Projects

Researchers analyzing manga visual grammar need clean panel images without text interference. PanelCleaner's mask-only export mode preserves original artwork while creating analysis-ready datasets—far superior to destructive editing.

4. Fan Remaster Projects

Community efforts to upscale or colorize classic manga require pristine source material. PanelCleaner extracts text layers separately, allowing colorists to work underneath without destructive editing. The cut-out text can even be preserved and reapplied over colored versions.

5. Personal Learning and Translation Practice

Aspiring translators often abandon practice due to cleaning overhead. PanelCleaner removes this barrier, letting learners focus on actual translation skills rather than Photoshop technique.

Step-by-Step Installation & Setup Guide

PanelCleaner offers multiple installation paths depending on your technical comfort and hardware. Here's how to get running:

Option 1: Pre-Built Binary (Recommended for Most Users)

Download the latest release for your platform from the releases page. Available for Windows (.exe) and Linux (.elf). Simply extract and run—no Python knowledge required.

Important limitation: Pre-built binaries do not support CUDA acceleration. For GPU speedups, use pip installation below.

Option 2: PyPI Installation (Python 3.10+)

This unlocks full features including CUDA support:

# Install with full GUI support
pip install pcleaner

# Or CLI-only for headless servers
pip install pcleaner-cli

For CUDA acceleration, first install PyTorch with your specific CUDA version from pytorch.org, then install pcleaner.

Option 3: Arch Linux (AUR) — Best CUDA Setup

The AUR package uses pipx isolation and allows PyTorch to auto-detect your CUDA version:

# Using yay AUR helper
yay -S panelcleaner

This provides pcleaner, pcleaner-gui commands, plus a desktop file for application menus.

Option 4: Flatpak (Containerized, No CUDA)

# Install from Flathub
flatpak install flathub io.github.voxelcubes.panelcleaner

Option 5: Docker (Cloud/Server Deployment)

# Build with modern buildx
docker buildx build -t pcleaner:v1 .

# Or legacy builder
docker image build -t pcleaner:v1 .

# Initialize with mounted working directory
docker run -it --name pcleaner -v $(pwd):/app pcleaner:v1

# Re-enter container later
docker start pcleaner
docker exec -it pcleaner bash

First Launch: All installation methods require downloading ~500MB of model data on first run. This is cached for future use.

REAL Code Examples from PanelCleaner

Let's examine actual usage patterns from the project's documentation, with detailed explanations:

Example 1: Basic Cleaning Command

# Clean individual images with automatic output directory creation
pcleaner clean image1.png image2.png image3.png

# The 'clean' subcommand is the primary operation
# It creates a 'cleaned/' directory alongside input files
# Output includes both cleaned images and mask visualizations

This is your bread-and-butter command. PanelCleaner accepts any number of image paths or directories, automatically organizing outputs. The default behavior applies cleaning directly to images, but as we'll see, mask-only mode is often more useful.

Example 2: Mask-Only Export for Professional Workflows

# Export only the transparent mask layer (-m shorthand)
pcleaner clean -m folder1 image1.png

# This produces PNG masks with alpha transparency
# Masks can be imported into Photoshop/GIMP as layers
# Non-destructive workflow: original art stays pristine

The -m flag is powerful. Instead of baking cleaned results, you get transparent masks representing detected text regions. Import these into your graphics editor as a layer mask, giving you manual control over final application. This is how professionals integrate AI assistance without surrendering creative control.

Example 3: Launching the GUI

# Launch graphical interface from command line
pcleaner gui

# Or directly if installed via certain methods
pcleaner-gui

# Fallback if PATH issues occur
python -m pcleaner

The GUI provides visual review of cleaning stages, interactive OCR editing, and profile management. For batch operations, CLI is faster; for fine-tuning individual pages, the GUI shines.

Example 4: Custom Profile Creation

# Generate new editable profile
pcleaner profile new my_profile_name_here

# Apply profile during cleaning
pcleaner clean -p my_profile_name_here image.png

# Cache intermediate mask visualizations for debugging
pcleaner clean --cache-masks image.png

Profiles are where PanelCleaner's sophistication becomes apparent. The default profile is optimized for ~1100x1600 pixel images. Here's a critical configuration snippet:

# Number of pixels to grow the mask by each step
# Smaller = more accurate but slower processing
# Larger = faster but may over-expand into artwork
mask_growth_step_pixels = 2

# Total expansion iterations
# Higher values create more generous masks
# Limited by reference box size to prevent runaway growth
mask_growth_steps = 11

Adjusting these two parameters dramatically changes behavior. The documentation shows a comparison: default settings versus mask_growth_step_pixels = 4 and mask_growth_steps = 4 produce visibly different mask aggressiveness. Use --cache-masks to visualize these intermediates and dial in settings for your specific manga style.

Example 5: OCR Text Extraction

# Extract Japanese text to file for translation assistance
pcleaner ocr myfolder --output-path=output.txt

# Combines MangaOCR (preferred for Japanese) 
# with Tesseract fallback for English/Japanese

This bridges cleaning and translation workflows. Extracted text can be fed into translation memory systems or shared with translators before cleaning even completes.

Advanced Usage & Best Practices

Profile Optimization for Your Source Material

The default profile targets standard web-release resolutions. For high-resolution magazine scans (3000+ pixels), proportionally increase size-dependent parameters. For low-quality web rips, reduce mask_growth_step_pixels to prevent over-expansion into degraded JPEG artifacts.

Hybrid Human-AI Workflow

PanelCleaner's 2-8% miss rate means human review remains essential. The most efficient workflow: run batch cleaning, quickly scan outputs for skipped bubbles (visually obvious as uncleaned text), manually handle those in your graphics editor. This 95/5 split of AI-to-human effort transforms productivity.

CUDA Environment Verification

If GPU acceleration seems absent, verify PyTorch CUDA availability:

import torch
print(torch.cuda.is_available())  # Must return True
print(torch.version.cuda)         # Should match your driver

Docker for Scalable Processing

The Docker deployment enables cloud-based processing pipelines. Mount network storage volumes, trigger cleaning via CI/CD webhooks, or distribute batch jobs across container orchestration platforms.

Mask Layer Compositing Trick

Export masks with -m, then in your editor: place mask as layer above original, invert mask selection, delete. This non-destructive approach lets you refine edges manually while preserving original data.

Comparison with Alternatives

Feature	PanelCleaner	Photoshop Content-Aware	Manual Cleaning	Generic Inpainting Tools
Manga-specific detection	✅ Trained on comic text	❌ Generic patterns	✅ Human judgment	❌ General purpose
Offline operation	✅ Fully local	⚠️ Cloud features optional	✅	Varies
Cost	✅ Free (GPL v3)	❌ Subscription	✅ Time only	Varies
Batch processing	✅ Native	⚠️ Actions possible	❌ Manual only	Varies
Conservative errors	✅ Skips uncertain	❌ May hallucinate fill	✅ Perfect judgment	❌ Often destructive
OCR integration	✅ Built-in	❌ Separate tools	❌ Separate tools	❌
Learning curve	Low-Moderate	High	Very High	Low
Speed (46 images)	Minutes (CUDA)	Hours to set up batch	Hours	Varies

PanelCleaner's unique advantage is domain-specific intelligence combined with conservative reliability. Generic tools don't understand manga bubble conventions; manual cleaning doesn't scale; Photoshop's AI lacks manga training data. PanelCleaner occupies the sweet spot of specialized enough to work brilliantly, general enough to handle varied manga styles.

FAQ: Your Burning Questions Answered

Q: What languages does PanelCleaner support for text cleaning? A: Japanese and English text cleaning are officially supported. Success rates may vary with other languages. Japanese OCR is fully supported; English OCR requires Tesseract installation.

Q: Can it handle colored speech bubbles? A: Currently, mask output supports grayscale only—white, black, or gray bubbles work optimally. Colored bubbles may produce incomplete results. The developers note this as a known limitation.

Q: How accurate is the AI detection really? A: From community testing, expect 92-98% of bubbles to be handled correctly depending on settings and source quality. The conservative approach means missed bubbles are left for manual cleaning rather than damaged by incorrect removal.

Q: Is my manga data sent to any cloud service? A: Absolutely not. After initial model download, all processing occurs locally on your machine. This is a core design principle for handling copyrighted and unpublished materials securely.

Q: Can I use this for commercial manga localization? A: PanelCleaner is GPL v3 licensed. Commercial use is permitted, but derivative works must also be open-sourced under GPL v3. Consult legal counsel for your specific use case; many publishers use GPL tools in proprietary pipelines through careful architectural separation.

Q: What file formats work with PanelCleaner? A: Input: JPEG, PNG, BMP, TIFF, JP2, DIB, WebP, PPM. Export adds PSD support. This covers virtually all manga distribution formats.

Q: How do I get help if something breaks? A: The FAQ documentation covers common issues. For platform-specific setup (especially macOS and exotic Linux distributions), community support is available through GitHub issues.

Conclusion: The Future of Manga Cleaning Is Here

PanelCleaner represents something rare in open-source tooling: a genuinely production-ready solution to a real, painful problem that affects thousands of people. It doesn't promise magic—it promises to eliminate the monotonous 80% of cleaning work so humans can focus on the nuanced 20% that actually requires judgment.

The manga localization ecosystem has been held back by cleaning bottlenecks for decades. Tools like PanelCleaner don't replace human cleaners; they elevate them from pixel-pushing technicians to quality controllers and artistic decision-makers. That's a future worth building toward.

If you're involved in manga translation, academic manga research, or any workflow requiring clean panel extraction, stop wasting hours on work that AI handles better. Install PanelCleaner today from the official repository, experiment with profiles for your specific material, and reclaim your time for creative work that actually matters.

The code is waiting. Your wrists will thank you.