/StyleMapGAN

Primary LanguagePythonOtherNOASSERTION

StyleMapGAN - Official PyTorch Implementation

StyleMapGAN: Exploiting Spatial Dimensions of Latent in GAN for Real-time Image Editing
Hyunsu Kim, Yunjey Choi, Junho Kim, Sungjoo Yoo, Youngjung Uh
In CVPR 2021.

Paper: https://arxiv.org/abs/2104.14754
5-minute video (CVPR): https://www.youtube.com/watch?v=7sJqjm1qazk
Demo video: https://youtu.be/qCapNyRA_Ng

Abstract: Generative adversarial networks (GANs) synthesize realistic images from random latent vectors. Although manipulating the latent vectors controls the synthesized outputs, editing real images with GANs suffers from i) time-consuming optimization for projecting real images to the latent vectors, ii) or inaccurate embedding through an encoder. We propose StyleMapGAN: the intermediate latent space has spatial dimensions, and a spatially variant modulation replaces AdaIN. It makes the embedding through an encoder more accurate than existing optimization-based methods while maintaining the properties of GANs. Experimental results demonstrate that our method significantly outperforms state-of-the-art models in various image manipulation tasks such as local editing and image interpolation. Last but not least, conventional editing methods on GANs are still valid on our StyleMapGAN. Source code is available at https://github.com/naver-ai/StyleMapGAN.

Demo

Youtube video Click the figure to watch the teaser video.

Interactive demo app Run demo in your local machine.

All test images are from CelebA-HQ, AFHQ, and LSUN.

python demo.py --ckpt expr/checkpoints/celeba_hq_256_8x8.pt --dataset celeba_hq

Installation

ubuntu gcc 7.4.0 CUDA CUDA-driver cudnn7 conda Python 3.6.12 pytorch 1.4.0

Clone this repository:

git clone https://github.com/naver-ai/StyleMapGAN.git
cd StyleMapGAN/

Install the dependencies:

conda create -y -n stylemapgan python=3.6.12
conda activate stylemapgan
./install.sh

Datasets and pre-trained networks

We provide a script to download datasets used in StyleMapGAN and the corresponding pre-trained networks. The datasets and network checkpoints will be downloaded and stored in the data and expr/checkpoints directories, respectively.

CelebA-HQ. To download the CelebA-HQ dataset and parse it, run the following commands:

# Download raw images and create LMDB datasets using them
# Additional files are also downloaded for local editing
bash download.sh create-lmdb-dataset celeba_hq

# Download the pretrained network (256x256) 
bash download.sh download-pretrained-network-256 celeba_hq # 20M-image-trained models
bash download.sh download-pretrained-network-256 celeba_hq_5M # 5M-image-trained models used in our paper for comparison with other baselines and for ablation studies.

# Download the pretrained network (1024x1024 image / 16x16 stylemap / Light version of Generator)
bash download.sh download-pretrained-network-1024 ffhq_16x16

AFHQ. For AFHQ, change above commands from 'celeba_hq' to 'afhq'.

Train network

Implemented using DistributedDataParallel.

# CelebA-HQ
python train.py --dataset celeba_hq --train_lmdb data/celeba_hq/LMDB_train --val_lmdb data/celeba_hq/LMDB_val

# AFHQ
python train.py --dataset afhq --train_lmdb data/afhq/LMDB_train --val_lmdb data/afhq/LMDB_val

# CelebA-HQ / 1024x1024 image / 16x16 stylemap / Light version of Generator
python train.py --size 1024 --latent_spatial_size 16 --small_generator --dataset celeba_hq --train_lmdb data/celeba_hq/LMDB_train --val_lmdb data/celeba_hq/LMDB_val 

Generate images

Reconstruction Results are saved to expr/reconstruction.

# CelebA-HQ
python generate.py --ckpt expr/checkpoints/celeba_hq_256_8x8.pt --mixing_type reconstruction --test_lmdb data/celeba_hq/LMDB_test

# AFHQ
python generate.py --ckpt expr/checkpoints/afhq_256_8x8.pt --mixing_type reconstruction --test_lmdb data/afhq/LMDB_test

W interpolation Results are saved to expr/w_interpolation.

# CelebA-HQ
python generate.py --ckpt expr/checkpoints/celeba_hq_256_8x8.pt --mixing_type w_interpolation --test_lmdb data/celeba_hq/LMDB_test

# AFHQ
python generate.py --ckpt expr/checkpoints/afhq_256_8x8.pt --mixing_type w_interpolation --test_lmdb data/afhq/LMDB_test

Local editing Results are saved to expr/local_editing. We pair images using a target semantic mask similarity. If you want to see details, please follow preprocessor/README.md.

# Using GroundTruth(GT) segmentation masks for CelebA-HQ dataset.
python generate.py --ckpt expr/checkpoints/celeba_hq_256_8x8.pt --mixing_type local_editing --test_lmdb data/celeba_hq/LMDB_test --local_editing_part nose

# Using half-and-half masks for AFHQ dataset.
python generate.py --ckpt expr/checkpoints/afhq_256_8x8.pt --mixing_type local_editing --test_lmdb data/afhq/LMDB_test

Unaligned transplantation Results are saved to expr/transplantation. It shows local transplantations examples of AFHQ. We recommend the demo code instead of this.

python generate.py --ckpt expr/checkpoints/afhq_256_8x8.pt --mixing_type transplantation --test_lmdb data/afhq/LMDB_test

Random Generation Results are saved to expr/random_generation. It shows random generation examples.

python generate.py --mixing_type random_generation --ckpt expr/checkpoints/celeba_hq_256_8x8.pt

Style Mixing Results are saved to expr/stylemixing. It shows style mixing examples.

python generate.py --mixing_type stylemixing --ckpt expr/checkpoints/celeba_hq_256_8x8.pt --test_lmdb data/celeba_hq/LMDB_test

Semantic Manipulation Results are saved to expr/semantic_manipulation. It shows local semantic manipulation examples.

python semantic_manipulation.py --ckpt expr/checkpoints/celeba_hq_256_8x8.pt --LMDB data/celeba_hq/LMDB --svm_train_iter 10000

Metrics

  • Reconstruction: LPIPS, MSE
  • W interpolation: FIDlerp
  • Generation: FID
  • Local editing: MSEsrc, MSEref, Detectability (Refer to CNNDetection)

If you want to see details, please follow metrics/README.md.

License

The source code, pre-trained models, and dataset are available under Creative Commons BY-NC 4.0 license by NAVER Corporation. You can use, copy, tranform and build upon the material for non-commercial purposes as long as you give appropriate credit by citing our paper, and indicate if changes were made.

For business inquiries, please contact clova-jobs@navercorp.com.
For technical and other inquires, please contact hyunsu1125.kim@navercorp.com.

Citation

If you find this work useful for your research, please cite our paper:

@inproceedings{kim2021stylemapgan,
  title={Exploiting Spatial Dimensions of Latent in GAN for Real-time Image Editing},
  author={Kim, Hyunsu and Choi, Yunjey and Kim, Junho and Yoo, Sungjoo and Uh, Youngjung},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  year={2021}
}

Related Projects

Model code starts from StyleGAN2 PyTorch unofficial code, which refers to StyleGAN2 official code. LPIPS, FID, and CNNDetection codes are used for evaluation. In semantic manipulation, we used StyleGAN pretrained network to get positive and negative samples by ranking. The demo code starts from Neural-Collage.