Project | Youtube | Paper

    Pytorch implementation of our method for high-resolution (e.g. 2048x1024) photorealistic image-to-image translation. It can be used for turning semantic label maps into photo-realistic images or synthesizing portraits from face label maps.

    High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs
    Ting-Chun Wang1, Ming-Yu Liu1, Jun-Yan Zhu2, Andrew Tao1, Jan Kautz1, Bryan Catanzaro1
    1NVIDIA Corporation, 2UC Berkeley
    In CVPR 2018.

    Image-to-image translation at 2k/1k resolution

    • Our label-to-streetview results

    - Interactive editing results

    - Additional streetview results

    • Label-to-face and interactive editing results

    • Our editing interface


    • Linux or macOS
    • Python 2 or 3
    • NVIDIA GPU (11G memory or larger) + CUDA cuDNN

    Getting Started


    pip install dominate
    • Clone this repo:
    git clone
    cd pix2pixHD


    • A few example Cityscapes test images are included in the datasets folder.
    • Please download the pre-trained Cityscapes model from here (google drive link), and put it under ./checkpoints/label2city_1024p/
    • Test the model (bash ./scripts/
    python --name label2city_1024p --netG local --ngf 32 --resize_or_crop none

    The test results will be saved to a html file here: ./results/label2city_1024p/test_latest/index.html.

    More example scripts can be found in the scripts directory.


    • We use the Cityscapes dataset. To train a model on the full dataset, please download it from the official website (registration required). After downloading, please put it under the datasets folder in the same way the example images are provided.


    • Train a model at 1024 x 512 resolution (bash ./scripts/
    python --name label2city_512p
    • To view training results, please checkout intermediate results in ./checkpoints/label2city_512p/web/index.html. If you have tensorflow installed, you can see tensorboard logs in ./checkpoints/label2city_512p/logs by adding --tf_log to the training scripts.

    Multi-GPU training

    • Train a model using multiple GPUs (bash ./scripts/
    python --name label2city_512p --batchSize 8 --gpu_ids 0,1,2,3,4,5,6,7

    Note: this is not tested and we trained our model using single GPU only. Please use at your own discretion.

    Training with Automatic Mixed Precision (AMP) for faster speed

    • To train with mixed precision support, please first install apex from:
    • You can then train the model by adding --fp16. For example,
    python -m torch.distributed.launch --name label2city_512p --fp16

    In our test case, it trains about 80% faster with AMP on a Volta machine.

    Training at full resolution

    • To train the images at full resolution (2048 x 1024) requires a GPU with 24G memory (bash ./scripts/, or 16G memory if using mixed precision (AMP).
    • If only GPUs with 12G memory are available, please use the 12G script (bash ./scripts/, which will crop the images during training. Performance is not guaranteed using this script.

    Training with your own dataset

    • If you want to train with your own dataset, please generate label maps which are one-channel whose pixel values correspond to the object labels (i.e. 0,1,...,N-1, where N is the number of labels). This is because we need to generate one-hot vectors from the label maps. Please also specity --label_nc N during both training and testing.
    • If your input is not a label map, please just specify --label_nc 0 which will directly use the RGB colors as input. The folders should then be named train_A, train_B instead of train_label, train_img, where the goal is to translate images from A to B.
    • If you don't have instance maps or don't want to use them, please specify --no_instance.
    • The default setting for preprocessing is scale_width, which will scale the width of all training images to opt.loadSize (1024) while keeping the aspect ratio. If you want a different setting, please change it by using the --resize_or_crop option. For example, scale_width_and_crop first resizes the image to have width opt.loadSize and then does random cropping of size (opt.fineSize, opt.fineSize). crop skips the resizing step and only performs random cropping. If you don't want any preprocessing, please specify none, which will do nothing other than making sure the image is divisible by 32.

    More Training/Test Details

    • Flags: see options/ and options/ for all the training flags; see options/ and options/ for all the test flags.
    • Instance map: we take in both label maps and instance maps as input. If you don't want to use instance maps, please specify the flag --no_instance.


    If you find this useful for your research, please use the following.

      title={High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs},
      author={Ting-Chun Wang and Ming-Yu Liu and Jun-Yan Zhu and Andrew Tao and Jan Kautz and Bryan Catanzaro},  
      booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},


    This code borrows heavily from pytorch-CycleGAN-and-pix2pix.


    🚀 Github 镜像仓库 🚀




    贡献者 9


    • Python 97.0 %
    • Shell 3.1 %