Project Page | Paper | Video | Talk | UI code

Real-Time User-Guided Image Colorization with Learned Deep Priors.
Richard Zhang*, Jun-Yan Zhu*, Phillip Isola, Xinyang Geng, Angela S. Lin, Tianhe Yu, and Alexei A. Efros.
In ACM Transactions on Graphics (SIGGRAPH 2017).

This is our PyTorch reimplementation for interactive image colorization, written by Richard Zhang and Jun-Yan Zhu.

This repository contains training usage. The original, official GitHub repo (with an interactive GUI, and originally Caffe backend) is here. The official repo has been updated to support PyTorch models on the backend, which can be trained in this repository.

• Linux or macOS
• Python 2 or 3
• CPU or NVIDIA GPU + CUDA CuDNN

• Install PyTorch 0.4+ and torchvision from https://pytorch.org and other dependencies (e.g., visdom and dominate). You can install all the dependencies by

pip install -r requirements.txt

• Clone this repo:

git clone https://github.com/richzhang/colorization-pytorch
cd colorization-pytorch

• Download the ILSVRC 2012 dataset and run the following script to prepare data python make_ilsvrc_dataset.py --in_path /PATH/TO/ILSVRC12. This will make symlinks into the training set, and divide the ILSVRC validation set into validation and test splits for colorization.

• Train a model: bash ./scripts/train_siggraph.sh. This is a 2 stage training process. First, the network is trained for automatic colorization using classification loss. Results are in ./checkpoints/siggraph_class. Then, the network is fine-tuned for interactive colorization using regression loss. Final results are in ./checkpoints/siggraph_reg2.

• To view training results and loss plots, run python -m visdom.server and click the URL https://localhost:8097. The following values are monitored:

  • G_CE is a cross-entropy loss between predicted color distribution and ground truth color.
  • G_entr is the entropy of the predicted distribution.
  • G_entr_hint is the entropy of the predicted distribution at points where a color hint is given.
  • G_L1_max is the L1 distance between the ground truth color and argmax of the predicted color distribution.
  • G_L1_mean is the L1 distance between the ground truth color and mean of the predicted color distribution.
  • G_L1_reg is the L1 distance between the ground truth color and the predicted color.
  • G_fake_real is the L1 distance between the predicted color and the ground truth color (in locations where a hint is given).
  • G_fake_hint is the L1 distance between the predicted color and the input hint color (in locations where a hint is given). It's a measure of how much the network "trusts" the input hint.
  • G_real_hint is the L1 distance between the ground truth color and the input hint color (in locations where a hint is given).

• Get a model. Either:

  • (1) download the pretrained model by running bash pretrained_models/download_siggraph_model.sh, which will give you a few models.
    • Original caffe weights [Recommended] ./checkpoints/siggraph_caffemodel/latest_net_G.pth is the original caffemodel weights, converted to PyTorch. It is recommended. Be sure to set --mask_cent 0 when running it.
    • Retrained model: ./checkpoints/siggraph_retrained/latest_net_G.pth. The model achieves better PSNR but performs qualitatively differently. Note that this repository is an approximate reimplementation of the siggraph paper.
  • (2) train your own model (as described in the section above), which will leave a model in ./checkpoints/siggraph_reg2/latest_net_G.pth

• Test the model on validation data:

  • python test.py --name siggraph_caffemodel --mask_cent 0 for original caffemodel weights
  • python test.py --name siggraph_retrained for retrained weights.
  • python test.py --name siggraph_reg2 if you retrained your own model The test results will be saved to an HTML file in ./results/[[NAME]]/latest_val/index.html. For each image in the validation set, it will test (1) automatic colorization, (2) interactive colorization with a few random hints, and (3) interactive colorization with lots of random hints.

• Test the model by making PSNR vs. the number of hints plot: python test_sweep.py --name [[NAME]] . This plot was used in Figure 6 of the paper. This test randomly reveals 6x6 color hint patches to the network and sees how accurate the colorization is with respect to the ground truth.

  

• Test the model interactively with the original official repository. Follow installation instructions in that repo and run python ideepcolor.py --backend pytorch --color_model [[PTH/TO/MODEL]] --dist_model [[PTH/TO/MODEL]].

If you use this code for your research, please cite our paper:

@article{zhang2017real,
  title={Real-Time User-Guided Image Colorization with Learned Deep Priors},
  author={Zhang, Richard and Zhu, Jun-Yan and Isola, Phillip and Geng, Xinyang and Lin, Angela S and Yu, Tianhe and Efros, Alexei A},
  journal={ACM Transactions on Graphics (TOG)},
  volume={9},
  number={4},
  year={2017},
  publisher={ACM}
}

This code borrows heavily from the pytorch-CycleGAN repository.