Disentangled Image Colorization via Global Anchors

Paper | Project Page | Online Demo

😊 This is the official PyTorch implementation of our colorization work DISCO, published in SIGGRAPH Asia 2022.

Conceptually, our disentangled colorization model consists of two components: (i) anchor color representation that predicts the global color anchors (i.e. location and probabilistic colors) as the color representation of the whole image; (ii) anchor-guided color generation that synthesizes the per-pixel colors deterministically by referring the specified anchor colors.

💼 Dependencies and Installation

• PyTorch >= 1.8.0
• CUDA >= 10.2
• Other required packages in requirements.txt

1. Clone this repository

git clone https://github.com/MenghanXia/DisentangledColorization
cd DisentangledColorization

2. Configure the environment

# create a new anaconda env
conda create -n DISCO python=3.8
source activate DISCO

# install pytortch
conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=10.2 -c pytorch

# install other packages
pip install requirements.txt

# create a anaconda env with required packages installed
conda env create -f environment.yaml

🎁 Checkpoints

⚡ Quick Inference

• Download Pre-trained Models: download a pretrained colorization model by sh ./checkpoints/disco_download.sh or from the tabulated links, and put it into the folder ./checkpoints.

• Prepare Testing Data: You can put the testing images in a folder, like ./data

• Test on Images: Run the inference script sh ./scripts/inferece.sh and the colorized images will be saved in ./test-anchor8. As default, the input image will be resized into 256x256 and colorized at this fixed resolution. Optional arguments includes:

  • --no_resize: colorize the image at the original input resolution (Not recommended because of unstable performance).
  • --diverse: generate diverse (three) colorization results.
  • --n_clusters: specify the number of color anchors (default=8).
  • --random_hint: use randomly scattered anchor locations.

Also, you can specify your own directories by runing the command below:

python ./main/colorizer/inference.py --checkpt [checkpoint path] --data [input dir] \
	--name [save name]

You are recommended to use the absolute paths as arguments, otherwise please note that running inference.py will redirect the current dir to ./main/colorizer. Note that, changing the random seed --seed may result in different colorization result because the clustering-based anchor location involves randomness.

💻 Training

• Download Pre-trained SPixelNet: download the pretrained SPixelNet-s16 and put it into a folder, like ./checkpoints.

• Prepare Data and Configuration: Official ImageNet and COCO dataset or your own dataset (named "disco") are supported. You need to specify the training arguments below:

  • --dataset: the dataset name ("imagenet", "coco", or "disco") that is required for dataloader construction.
  • --data_dir: the dataset location. If it is not official ImageNet or COCO, please organize the dataset folder as below so as to use our dataset configuration "disco" directly:

   ├─dataset
   |   ├─train
   |   |   └xxx.png
   |   |   └xxx.png
   |   ├─val
   |   |   └xxx.png
   |   |   └xxx.png

  • ckpt_dir: the directory of any pre-trained models required by the training, e.g. the pre-trained SPixelNet.
  • save_dir: the directory to save the training meta data and checkpoints.

• Train the Model: Again, you are recommended to use the absolute paths as arguments to avoid accident.

   sh scripts/train_yours_ddp.sh
  

📐 Evaluation

We provide the python implementation of the colorization evaluation metrics HERE, and the corresponding running scripts are attached.

# fidelity metrics: PSNR, SSIM, LPIPS
sh run_fidelity.sh

# perceptual quality: FID, IS, colorfulness
sh run_perception.sh

Acknowledgement

Part of our codes are taken from from SpixelFCN, iDeepColor, and DETR. Thanks for their awesome works.

Citation

If any part of our paper and code is helpful to your work, please generously cite with:

@article{XiaHWW22,
	author   = {Menghan Xia and Wenbo Hu and Tien-Tsin Wong and Jue Wang},
	title    = {Disentangled Image Colorization via Global Anchors},
	journal  = {ACM Transactions on Graphics (TOG)},
	volume   = {41},
	number   = {6},
	pages    = {204:1--204:13},
	year = {2022}
}