ReOldify

A community-driven continuation of DeOldify — Deep Learning based colorization and restoration of old images and video.

DeOldify was created by Jason Antic in 2018 and archived in October 2024. ReOldify picks up where it left off, with a modernized stack, a new state-of-the-art colorization engine (DDColor), and continued development.

---

What's New in ReOldify

• DDColor engine (ICCV 2023): State-of-the-art colorization with significantly richer, more accurate colors
• Modernized stack: Python 3.10+, PyTorch 2.x, updated dependencies
• CLI tool: reoldify colorize photo.jpg --model ddcolor --cpu
• REST API: FastAPI server to colorize via HTTP
• Docker: One-command deployment with Docker Compose
• CI/CD: GitHub Actions for linting and testing
• Modern packaging: pyproject.toml with proper dependency management

Quick Start

Installation

git clone https://github.com/pattex67/ReOldify.git
cd ReOldify
pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu
pip install -e .

CLI Usage

# Colorize with DDColor (best quality, weights auto-downloaded)
reoldify colorize photo.jpg --model ddcolor --cpu

# DDColor tiny (faster, lighter)
reoldify colorize photo.jpg --model ddcolor-tiny --cpu

# Legacy DeOldify models
reoldify colorize photo.jpg --model artistic --cpu
reoldify colorize photo.jpg --model stable --cpu

# Colorize a video
reoldify colorize video.mp4 --cpu

# Adjust render quality (higher = better but slower, default: 35)
reoldify colorize photo.jpg --model ddcolor --render-factor 45

# Specify output path
reoldify colorize photo.jpg -o colorized_photo.jpg --model ddcolor --cpu

Python API

from deoldify import device as device_settings
from deoldify.device_id import DeviceId
device_settings.set(DeviceId.CPU)

# DDColor (recommended)
from deoldify.visualize import get_ddcolor_image_colorizer
colorizer = get_ddcolor_image_colorizer(model_name="ddcolor")
result = colorizer.get_transformed_image("old_photo.jpg", render_factor=35, watermarked=False)
result.save("colorized.jpg")

# Legacy DeOldify
from deoldify.visualize import get_image_colorizer
colorizer = get_image_colorizer(artistic=True)
result = colorizer.get_transformed_image("old_photo.jpg", render_factor=35, watermarked=False)
result.save("colorized.jpg")

REST API

pip install fastapi uvicorn python-multipart
python -m deoldify.api

Then colorize images via HTTP:

curl -X POST http://localhost:8000/colorize \
  -F "file=@old_photo.jpg" \
  -F "model=ddcolor" \
  -o colorized.png

Docker

docker compose up --build

Available Models

Model	Engine	CLI flag	Size	Best for
DDColor	DDColor (2023)	--model ddcolor	912 MB	Best quality, general use
DDColor Tiny	DDColor (2023)	--model ddcolor-tiny	220 MB	Fast / CPU inference
Artistic	DeOldify (2019)	--model artistic	243 MB	Vibrant colors (legacy)
Stable	DeOldify (2019)	--model stable	243 MB	Portraits (legacy)
Video	DeOldify (2019)	--model video	243 MB	Video colorization

DDColor weights are auto-downloaded from HuggingFace on first use.

DeOldify weights must be downloaded manually to the models/ folder:

Model	Download
Artistic	ColorizeArtistic_gen.pth
Stable	ColorizeStable_gen.pth
Video	ColorizeVideo_gen.pth

Technical Background

DDColor (ICCV 2023)

DDColor uses a dual-decoder architecture: a pixel decoder restores spatial resolution while a query-based color decoder refines learnable color tokens via cross-attention. This produces richer, more accurate colors with fewer artifacts than previous approaches. Uses ConvNeXt as backbone encoder.

DeOldify (2019)

DeOldify uses the NoGAN training technique, combining GAN training benefits with minimal direct GAN training time. Uses a U-Net architecture with ResNet34 (artistic) or ResNet101 (stable) backbones.

Hardware Requirements

• Inference: GPU recommended for speed. CPU works fine (especially with ddcolor-tiny)
• Training: GPU with 11GB+ VRAM

Credits

• DeOldify by Jason Antic (MIT License)
• DDColor by Xiaoyang Kang et al. (Apache-2.0 License)

Contributing

Contributions are welcome! Areas where help is appreciated:

• Improving color accuracy
• ONNX/TensorRT export for fast inference
• Adding batch processing support
• Video support with DDColor

License

MIT License — see LICENSE for details.

DDColor architecture code is derived from the DDColor project under the Apache-2.0 license.