Manga Light Colorizer — ONNX Inference
Standalone inference script for the Manga Light Colorizer model.
Gallery
The following gallery uses the same source images as the manga-colorization-v2 project to facilitate direct comparison between models.
Comparison between input (left) and colorized output (right):
| Input (BW) | Colorized Output |
|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
Quick Start
# Install dependencies
pip install -r requirements.txt
# Single image
python inference.py --input input/bw1.jpg
# All images in a folder
python inference.py --input input/
# Custom output folder
python inference.py --input input/ --output_dir output/
# Custom inference resolution
python inference.py --input input/ --infer-size 1024
Arguments
| Argument | Required | Default | Description |
|---|---|---|---|
--input |
Yes | - | Input grayscale image or folder |
--onnx-model |
No | models/v6_generator.onnx |
Generator ONNX model path |
--sam-onnx |
No | models/v6_sam_encoder.onnx |
SAM 2.1 encoder ONNX path |
--output_dir |
No | ./output/ |
Output folder for colorized images |
--infer-size |
No | 768 |
Inference resolution (square) |
--ort-device |
No | cpu |
ONNX Runtime device (cpu or cuda) |
Model Information
- Architecture: FastViT-SA36 Encoder + DualSemanticSAM Guide + UNet V6 Decoder
- Training Resolution: 512×512 pixels
- Current Inference Resolution: 768×768 pixels (default)
- Output: Resized back to original input resolution
Important: Resolution Notice
The model was trained at 512×512 pixels. Inference currently runs at 768×768 pixels by default.
More the inference resolution differs from 512×512, the less faithful the colors will be.
For best results, use the training resolution:
# Best color accuracy, but lower resolution — matches training resolution python inference.py --input input/ --infer-size 512 # Default (good quality) python inference.py --input input/ # Higher resolution (may reduce color accuracy) python inference.py --input input/ --infer-size 1024
Pipeline
Input (grayscale) → Resize to infer-size → SAM 2.1 (zeros) → Generator ONNX → Resize to original
Requirements
- Python 3.10+
- onnxruntime
- numpy
- opencv-python
See requirements.txt for full list.
License
Model Weights
Licensed under CC BY-NC-SA 4.0 (Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International).
You may:
- Share — copy and redistribute the material in any medium or format
- Adapt — remix, transform, and build upon the material
Under the following terms:
- Attribution — You must give appropriate credit
- NonCommercial — You may not use the material for commercial purposes
- ShareAlike — If you remix, transform, or build upon the material, you must distribute your contributions under the same license
See: https://creativecommons.org/licenses/by-nc-sa/4.0/
Inference Code
Licensed under GNU General Public License v3 (GPL-3.0).
You may use, modify, and distribute this code under the terms of the GPL-3.0 license.