re-DDColor: Enhanced Image Colorization

A revised and improved version of DDColor, offering additional features and optimizations for image colorization.

This project is developed as part of my Bachelor of Computer Engineering with Honours Final Year Project (2025).

Overview

DDColor is a state-of-the-art automatic image colorization framework proposed in the paper "DDColor: Towards Photo-Realistic Image Colorization via Dual Decoders" (ICCV 2023). This revised version maintains the core functionality while adding new features, optimizing performance, and improving user interfaces.

Example Results

Below is a comparison of the input black-and-white images and the colorized outputs generated by re-DDColor:

Before (Black & White)	After (Colorized)

You can try colorizing your own images using the methods described in the Usage section.

New Features in This Version

Performance Improvements

• Added support for Apple Metal Performance Shaders (MPS) for better performance on Mac devices
• Optimized memory usage for improved processing on devices with limited resources
• Enhanced inference pipeline for faster colorization

Interface Enhancements

• Gradio Web UI: Added interactive web interface for easy colorization with before/after slider comparison
• Flask API: Implemented a REST API for integration with other applications
• Video Processing: Extended support for colorizing videos frame by frame
• Improved CLI: Enhanced command-line interface with more options and better feedback

Technical Improvements

• Better cross-platform compatibility (Windows, macOS, Linux)
• Streamlined installation process with clearer dependencies
• Additional model configuration options
• Enhanced error handling and recovery
• Comprehensive documentation

Installation

# Clone the repository
git clone https://github.com/yourusername/re-DDColor.git
cd re-DDColor

# Install dependencies
pip install -r requirements.txt

# Download pre-trained models (automatically done on first run)
python pretraindownload.py

Docker Implementation

For the easiest and most consistent experience across different platforms, you can use Docker:

# Build the Docker image
docker build -t ddcolor-app .

# Run the container
docker run -p 8501:8501 ddcolor-app

# Access the web application
# Open http://localhost:8501 in your browser

Advantages of using Docker:

• No need to install dependencies manually
• Works consistently across Windows, macOS, and Linux
• Model is downloaded automatically during the build process
• All required libraries are pre-configured

Requirements:

• Docker installed on your system (Get Docker)
• At least 2GB of free disk space

Usage

Web Interface (Gradio)

The easiest way to use re-DDColor is through the Gradio web interface:

python gradio_app.py

This will start a local web server with an interactive UI where you can:

• Upload black and white images
• View colorization results with a before/after slider
• Download the colorized images

REST API (Flask)

For integration with other applications:

python flaskapi.py

The API will be available at http://localhost:5000/ with the following endpoints:

• POST /colorize: Upload an image for colorization
• GET /models: List available models

Command Line

For batch processing or scripting:

# Colorize a single image
python predict.py --input path/to/image.jpg --output path/to/output.jpg

# Colorize all images in a directory
python predict.py --input path/to/input_dir --output path/to/output_dir

# Specify a different model
python predict.py --input image.jpg --output output.jpg --model ddcolor_artistic.pth

Video Colorization

To colorize videos:

python video_converter.py --input path/to/video.mp4 --output path/to/colorized_video.mp4

Available Models

re-DDColor includes several pre-trained models as described in MODEL_ZOO.md:

Model	Description	Best For
ddcolor_modelscope.pth (default)	DDColor-L with cleaned data	General use, best quality
ddcolor_paper.pth	Original paper model	Paper reproduction
ddcolor_artistic.pth	Trained with artistic images	Creative colorization, fewer artifacts
ddcolor_paper_tiny.pth	Lightweight model	Resource-constrained devices

Technical Details

Architecture

re-DDColor uses a dual-decoder architecture:

1. Encoder: ConvNeXt-based (tiny or large variants)
2. Decoder: Two options:
   • MultiScaleColorDecoder: Transformer-based decoder with multi-scale features
   • SingleColorDecoder: Simpler decoder for faster inference
3. Color Processing: Lab color space for accurate colorization

System Requirements

• Python 3.7+
• PyTorch 1.7+
• CUDA-compatible GPU (recommended) or Apple Silicon
• 4GB+ VRAM for large models, 2GB+ for tiny models

Training Your Own Model

To train a custom model:

# Edit the configuration
# Modify options/train/train_ddcolor.yml with your dataset paths and parameters

# Start training
python basicsr/train.py -opt options/train/train_ddcolor.yml

See the training configuration file for details on available options including loss functions, optimizers, and data augmentation.

Citation

If you use this code in your research, please cite the original DDColor paper:

@inproceedings{kang2023ddcolor,
  title={DDColor: Towards Photo-Realistic Image Colorization via Dual Decoders},
  author={Kang, Xiaoyang and Yang, Tao and Ouyang, Wenqi and Ren, Peiran and Li, Lingzhi and Xie, Xuansong},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  year={2023}
}

Acknowledgements

This work is a revised version of the original DDColor project by Xiaoyang Kang et al. I thank the original authors for their outstanding contribution to the field of image colorization.

License

This project is licensed under the terms of the LICENSE file included in the repository.