AI-powered restoration of physical Weekly Shonen Jump manga scans
schillyCL is a fine-tuned image restoration model built on Real-ESRGAN, trained to clean physical WSJ magazine scans by removing paper grain, yellowing, halftone artifacts, and restoring black ink lines to their original clarity.
| Before | After |
|---|---|
 |
 |
| Before | After |
|---|---|
 |
 |
| Before | After |
|---|---|
 |
 |
| Before | After |
|---|---|
 |
 |
- Removes paper grain and yellowing from physical magazine scans
- Reduces halftone dot patterns while preserving fine detail
- Restores black ink lines without over-smoothing
- Tile-based inference with Gaussian blending giving no visible seams
- Optional bilateral filter pre-processing and levels post-processing
- Outputs at 2200px height by default
schillyCL is based on RRDBNet (Real-ESRGAN architecture) fine-tuned on a dataset of matched physical WSJ scan / digital pairs spanning 20+ chapters across multiple manga titles.
Training highlights:
- Real scan/digital pairs aligned via ORB feature matching
- Custom WeightedL1Loss punishing incorrect black pixel output 3x harder than standard L1
- Scale 1 restoration; no upscaling, pure quality improvement
- Trained on RTX 3060 12GB
Model weights are not included in this repository.
Download weights: coming soon
git clone https://github.com/schilly-stack/schillyCL
cd schillyCL
pip install -r requirements.txtYou will also need to install Real-ESRGAN and BasicSR:
pip install basicsr realesrganpython inference.pyReads from data/test/ and outputs to data/test_output/.
python inference_filtered.pyApplies bilateral filter before the model and levels adjustment after.
If you have matched scan/digital pairs:
# 1. Rename files to sequential numbers per chapter
python scripts/prepare/rename.py
# 2. Align scans to digitals and generate crops
python scripts/prepare/align_and_crop.py
# 3. Clean up unmatched crops
python scripts/prepare/cleanup_orphans.pyIf you only have digitals and want to use synthetic degradation:
python scripts/prepare/degrade.pyTraining requires Real-ESRGAN installed locally and a matched dataset of scan/digital pairs or synthetically degraded images.
See scripts/prepare/ for dataset preparation tools.
Training is configured via options/manga_finetune.yml in your Real-ESRGAN installation.
Key training parameters used:
- Architecture: RRDBNet (num_block: 23, num_feat: 64)
- Loss: WeightedL1Loss (dark_weight: 3.0, dark_threshold: 0.3)
- Patch size: 512×512
- Optimizer: Adam (lr: 1e-4)
- Total iterations: 50,000+
Built on top of Real-ESRGAN by Xintao Wang et al.
Training data sourced from Weekly Shonen Jump physical magazine scans.
MIT License — model weights subject to separate terms when released.