Home

Awesome

Few-shot Semantic Image Synthesis Using StyleGAN Prior

The extended version is available <a href="https://github.com/endo-yuki-t/StyleGANSparseControl">here</a>.

<a href="https://arxiv.org/abs/2103.14877"><img src="https://img.shields.io/badge/arXiv-2103.14877-b31b1b.svg"></a> <a href="https://opensource.org/licenses/MIT"><img src="https://img.shields.io/badge/License-MIT-yellow.svg"></a>

<p align="center"> <img src="docs/teaser.jpg" width="800px"/> <br> Our method can synthesize photorealistic images from dense or sparse semantic annotations using a few training pairs and a pre-trained StyleGAN. </p>

Prerequisites

  1. Python3
  2. PyTorch

Preparation

Download and decompress <a href="https://drive.google.com/file/d/1s-AS7WRUbL3MzEALxM8y4_XO3n3panxH/view?usp=sharing">the file containing StyleGAN pre-trained models</a> and put the "pretrained_models" directory in the parent directory.

Inference with our pre-trained models

  1. Download and decompress <a href="https://drive.google.com/file/d/1INoO3WpEY8hkx7LyOdiOX-dPBNEO4gSs/view?usp=sharing">the file containing our pretrained encoders</a> and put the "results" directory in the parent directory.
  2. For example, our results for celebaMaskHQ in a one-shot setting can be generated as follows:
python scripts/inference.py --exp_dir=results/celebaMaskHQ_oneshot --checkpoint_path=results/celebaMaskHQ_oneshot/checkpoints/iteration_100000.pt --data_path=./data/CelebAMask-HQ/test/labels/ --couple_outputs --latent_mask=8,9,10,11,12,13,14,15,16,17

Inference results are generated in results/celebaMaskHQ_oneshot. If you use other datasets, please specify --exp_dir, --checkpoint_path, and --data_path appropriately.

<p align="center"> <img src="docs/results.jpg" width="800px"/> </p>

Training

For each dataset, you can train an encoder as follows:

python scripts/train.py --exp_dir=[result_dir] --dataset_type=celebs_seg_to_face --stylegan_weights pretrained_models/stylegan2-ffhq-config-f.pt --start_from_latent_avg --label_nc 19 --input_nc 19

python scripts/train.py --exp_dir=[result_dir] --dataset_type=celebs_landmark_to_face --stylegan_weights pretrained_models/stylegan2-ffhq-config-f.pt --start_from_latent_avg --label_nc 71 --input_nc 71 --sparse_labeling
<p align="center"> <img src="docs/int_outputs.jpg" width="800px"/> <br> Intermediate training outputs with the StyleGAN pre-trained with the CelebA-HQ dataset. It can be seen that the layouts of the bottom-row images reconstructed from the middle-row pseudo semantic masks gradually become close to those of the top-row StyleGAN samples as the training iterations increase. </p> 
python scripts/train.py --exp_dir=[result_dir] --dataset_type=lsunchurch_seg_to_img --stylegan_weights pretrained_models/stylegan2-church-config-f.pt --style_num 14 --start_from_latent_avg --label_nc 151 --input_nc 151

python scripts/train.py --exp_dir=[result_dir] --dataset_type=lsuncar_seg_to_img --stylegan_weights pretrained_models/stylegan2-car-config-f.pt --style_num 16 --start_from_latent_avg --label_nc 5 --input_nc 5

python scripts/train.py --exp_dir=[result_dir] --dataset_type=lsuncat_scribble_to_img --stylegan_weights pretrained_models/stylegan2-cat-config-f.pt --style_num 14 --start_from_latent_avg --label_nc 9 --input_nc 9 --sparse_labeling

python scripts/train.py --exp_dir=[result_dir] --dataset_type=ukiyo-e_scribble_to_img --stylegan_weights pretrained_models/ukiyoe-256-slim-diffAug-002789.pt --style_num 14 --channel_multiplier 1 --start_from_latent_avg --label_nc 8 --input_nc 8 --sparse_labeling

python scripts/train.py --exp_dir=[result_dir] --dataset_type=anime_cross_to_img --stylegan_weights pretrained_models/2020-01-11-skylion-stylegan2-animeportraits-networksnapshot-024664.pt --style_num 16 --start_from_latent_avg --label_nc 2 --input_nc 2 --sparse_labeling

Using StyleGAN samples as few-shot training data

  1. Run the following script:
python scripts/generate_stylegan_samples.py --exp_dir=[result_dir] --stylegan_weights ./pretrained_models/stylegan2-ffhq-config-f.pt --style_num 18 --channel_multiplier 2

Then a StyleGAN image (*.png) and a corresponding latent code (*.pt) are obtained in [result_dir]/data/images and [result_dir]/checkpoints.

  1. Manually annotate the generated image in [result_dir]/data/images and save the annotated mask in [result_dir]/data/labels.

  2. Edit ./config/data_configs.py and ./config/paths_config.py appropriately to use the annotated pairs as a training set.

  3. Run a training command above with appropriate options.

Citation

Please cite our paper if you find the code useful:

@article{endo2021fewshotsmis,
  title = {Few-shot Semantic Image Synthesis Using StyleGAN Prior},
  author = {Yuki Endo and Yoshihiro Kanamori},
  journal   = {CoRR},
  volume    = {abs/2103.14877},
  year      = {2021}
}

Acknowledgements

This code heavily borrows from the pixel2style2pixel repository.