Towards Vivid and Diverse Image Colorization with Generative Color Prior, ICCV 2021

Yanze Wu, Xintao Wang, Yu Li, Honglun Zhang, Xun Zhao, Ying Shan
ARC Lab, Tencent PCG

---

Installation

Core dependencies

• CUDA >= 10.0 (test on 10.0, 10.2, and 11.1)
• gcc >= 7.3
• pytorch >= 1.6 (test on 1.6, 1.7.1, and 1.9.0)
• python3 (test on 3.8)
• yacs

Install DCN package (only required by torchvision < 0.9.0)

# you can skip this step if you are using torchvision >= 0.9.0
cd ops
python setup.py develop

Download the Pre-trained Models

Download the pretrained models from [Google Drive ] and put them into the assets folder. If you want to reproduce or get the results reported in our ICCV 2021 paper for academic purpose, you can check model zoo. Otherwise, you just need to use the default options, which is our best model.

Inference

Test in the wild images

1. Predict ImageNet label (0-999) for these images

   1. install awesome timm, pip install timm
   2. use a SOTA classification model from timm to predict the labels

      python predict_imagenet_label.py testcase_in_the_wild --model beit_large_patch16_512 --pretrained

      here testcase_in_the_wild folder has the images you want to test
   3. you will get the label map in assets/predicted_label_for_user_image.txt

2. Inference colorization

   python main.py --expname inference_in_the_wild --test_folder testcase_in_the_wild DATA.FULL_RES_OUTPUT True

   options:
       --expname: the results will be saved in results/{expname}-{time} folder
       --inference_in_the_wild: contains the images you want to test
       --bs: batch size
       DATA.FULL_RES_OUTPUT: True or False. If set to True
                             the full resolution results will be saved in results/{expname}-{time}/full_resolution_results
                             batch size should be 1 if this flag is set to True
       DATA.CENTER_CROP: whether to center crop the input images.
                         This flag and DATA.FULL_RES_OUTPUT flag can not be both set to True

3. If everything goes well, you will get the results similar as visual_results.md.

Test images from ImageNet val set

• The most worry-free way is to make sure the images' names are consistent with the official ImageNet name. Because we provide the GT ImageNet labels in imagenet_val_label_map.txt. You can check testcase_imagenet for examples.

• Also, the test images should better be color images rather than grayscale images. If you want to test on grayscale images, please read the read_data_from_dataiter function from base_solver.py, prepare the grayscale images following the pipeline and hack the related code.

• Inference

  python main.py --expname inference_imagenet --test_folder testcase_imagenet DATA.FULL_RES_OUTPUT False DATA.CENTER_CROP False

• If everything goes well, you will get the following quantitative results on the full 50,000 ImageNet validation images (yes, FID is better than the number reported in our ICCV 2021 paper)

  eval@256x256	FID↓	Colorfulness↑	ΔColorfulness↓
  w/o center crop	1.325	34.89	3.45
  w/ center crop	1.262	34.74	4.12

Diverse Colorization

You can achieve diverse colorization by:

1. adding random noise to the latent code

   python main.py --expname inference_random_diverse_color --latent_direction -1 --test_folder testcase_in_the_wild/testcase19_from_eccv2022_bigcolor.png DATA.FULL_RES_OUTPUT True

   Input	Diverse colorization	Diverse colorization	Diverse colorization

2. walking through the interpretable latent space

   we use the method described in Unsupervised Discovery of Interpretable Directions in the GAN Latent Space to find the interpretable directions for BigGAN. By setting latent_direction to the color-relevant direction (e.g., 4, 6, 23) we found, you can achieve more controllable diverse colorization.

   ps: we also provide the checkpoint (i.e., the full direction matrix), so you can find more color-relevant directions by yourself.

   # get the label first
   python predict_imagenet_label.py testcase_diverse --model beit_large_patch16_512 --pretrained
   # the 1st direction we found
   python main.py --expname inference_diverse_color_dir4 --latent_direction 4 --test_folder testcase_diverse/test_diverse_inp0.png
   # the 2nd direction we found
   python main.py --expname inference_diverse_color_dir6 --latent_direction 6 --test_folder testcase_diverse/test_diverse_inp1.png
   python main.py --expname inference_diverse_color_dir6 --latent_direction 6 --test_folder testcase_diverse/test_diverse_inp2.png
   # the 3rd direction we found
   python main.py --expname inference_diverse_color_dir23 --latent_direction 23 --test_folder testcase_diverse/test_diverse_inp3.png

   Input	Diverse colorization	Diverse colorization	Diverse colorization

3. changing the category

   you can modify the assets/predicted_label_for_user_image.txt

TODO

• add colab demo

Citation

If you find this project useful for your research, please consider citing our paper:

@inproceedings{wu2021towards,
  title={Towards vivid and diverse image colorization with generative color prior},
  author={Wu, Yanze and Wang, Xintao and Li, Yu and Zhang, Honglun and Zhao, Xun and Shan, Ying},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  year={2021}
}

Acknowedgement

This project borrows codes from CoCosNet, DGP, and BigGAN-PyTorch. DCN code from early version of mmcv. predict_imagenet_label.py from timm. Thanks the authors for sharing their awesome projects.