Tensorflow Implementation of Image2Stylegan++

This repository is Tensorflow Implementation of Image2StyleGAN++ proposed by Image2StyleGAN++: How to Edit the Embedded Images?. The StyleGAN structure is mainly based on taki0112 StyleGAN-Tensorflow architecture. In this repository, we only implement reconstruction, crossover, inpainting and style transfer(Image2StyleGAN).

Installation

• Clone this repository

  git clone https://github.com/Rayhchs/Image2Styleganv2-Implmentation.git

• Setup environment

  • python3.8

  pip install -r requirements.txt

• Download stylegan checkpoint from here.

Or you want to convert your pretrained checkpoint, you can follow:

Checkpoint Conversion

This repository converts NVlabs pretrained model into taki0112 architecture using stylegan convert architecture. Or you can download the converted pretrained model from here.

Usage

sh run.sh

That's it. Image results will be saved to ./result.

Important Arguments

Argument	Explanation
phase	Which process to be implement including reconstruction, crossover, inpainting, style_transfer
img_a	Filename of master image
img_b	Filename of slave image
test_folder	Folder of img_a and img_b
mask	Mask for inpainting and crossover
epoch_w	Iteration for w+ space optimization
epoch_n	Iteration for noise space optimization

Advanced Usage

This repository uses totally different lambda value from papers. If you are interested, you can modify args.lambda_mse, args.lambda_p, args.lambda_mse1, args.lambda_mse2 in main.py for your implementation.

Results

Phase	Master image	Slave image	Result
Reconstruction
Reconstruction
Crossover mask=[350, 650, 650, 850]
Crossover mask=[350, 650, 650, 850]
Inpainting mask=[350, 650, 650, 850]
Inpainting mask=[350, 650, 650, 850]
Style Transfer
Style Transfer
Style Transfer
Style Transfer
Style Transfer
Style Transfer

Acknowledge

• Tensorflow StyleGAN structure is borrowed from taki0112 StyleGAN-Tensorflow.

• Pretrained model is borrowed from NVlabs.

• Model conversion is based on stylegan convert architecture

Thanks for amazing works!

References

• StyleGAN

@article{DBLP:journals/corr/abs-1812-04948,
  author       = {Tero Karras and
                  Samuli Laine and
                  Timo Aila},
  title        = {A Style-Based Generator Architecture for Generative Adversarial Networks},
  journal      = {CoRR},
  volume       = {abs/1812.04948},
  year         = {2018},
  url          = {http://arxiv.org/abs/1812.04948},
  eprinttype    = {arXiv},
  eprint       = {1812.04948},
  timestamp    = {Tue, 01 Jan 2019 15:01:25 +0100},
  biburl       = {https://dblp.org/rec/journals/corr/abs-1812-04948.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}

• Image2StyleGAN

@article{DBLP:journals/corr/abs-1904-03189,
  author       = {Rameen Abdal and
                  Yipeng Qin and
                  Peter Wonka},
  title        = {Image2StyleGAN: How to Embed Images Into the StyleGAN Latent Space?},
  journal      = {CoRR},
  volume       = {abs/1904.03189},
  year         = {2019},
  url          = {http://arxiv.org/abs/1904.03189},
  eprinttype    = {arXiv},
  eprint       = {1904.03189},
  timestamp    = {Sat, 23 Jan 2021 01:19:25 +0100},
  biburl       = {https://dblp.org/rec/journals/corr/abs-1904-03189.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}

• Image2StyleGAN++

@article{DBLP:journals/corr/abs-1911-11544,
  author       = {Rameen Abdal and
                  Yipeng Qin and
                  Peter Wonka},
  title        = {Image2StyleGAN++: How to Edit the Embedded Images?},
  journal      = {CoRR},
  volume       = {abs/1911.11544},
  year         = {2019},
  url          = {http://arxiv.org/abs/1911.11544},
  eprinttype    = {arXiv},
  eprint       = {1911.11544},
  timestamp    = {Mon, 26 Jun 2023 20:49:40 +0200},
  biburl       = {https://dblp.org/rec/journals/corr/abs-1911-11544.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}