Skip to content

Latest commit

 

History

History
131 lines (101 loc) · 4.14 KB

styleganv2.md

File metadata and controls

131 lines (101 loc) · 4.14 KB
Raw

StyleGAN V2

StyleGAN V2 introduction

The task of StyleGAN V2 is image generation. Given a vector of a specific length, generate the image corresponding to the vector. It is an upgraded version of StyleGAN, which solves the problem of artifacts generated by StyleGAN.

StyleGAN V2 can mix multi-level style vectors. Its core is adaptive style decoupling.

Compared with StyleGAN, its main improvement is:

  • The quality of the generated image is significantly better (higher FID score, fewer artifacts)
  • Propose a new method to replace progressive training, with more perfect details such as teeth and eyes
  • Style mixing improved
  • Smoother interpolation
  • Train faster

How to use

Generate

The user can generate different results by replacing the value of the seed or removing the seed. Use the following command to generate images：

cd applications/
python -u tools/styleganv2.py \
       --output_path <DIRECTORY TO STORE OUTPUT IMAGE> \
       --weight_path <YOUR PRETRAINED MODEL PATH> \
       --model_type ffhq-config-f \
       --seed 233 \
       --size 1024 \
       --style_dim 512 \
       --n_mlp 8 \
       --channel_multiplier 2 \
       --n_row 3 \
       --n_col 5 \
       --cpu

params:

  • output_path: the directory where the generated images are stored
  • weight_path: pretrained model path
  • model_type: inner model type in PaddleGAN. If you use an existing model type, weight_path will have no effect. Currently available: ffhq-config-f, animeface-512
  • seed: random number seed
  • size: model parameters, output image resolution
  • style_dim: model parameters, dimensions of style z
  • n_mlp: model parameters, the number of multi-layer perception layers for style z
  • channel_multiplier: model parameters, channel product, affect model size and the quality of generated pictures
  • n_row: the number of rows of the sampled image
  • n_col: the number of columns of the sampled picture
  • cpu: whether to use cpu inference, if not, please remove it from the command

Train

prepapre datasets

you can get ffhq dataset from here

for convenient, we provide images256x256.tar

The structure of stylegan data is as following:

  PaddleGAN
    ├── data
        ├── ffhq
              ├──images1024x1024
                    ├── 00000.png
                    ├── 00001.png
                    ├── 00002.png
                    ├── 00003.png
                    ├── 00004.png
               ├──images256x256
                    ├── 00000.png
                    ├── 00001.png
                    ├── 00002.png
                    ├── 00003.png
                    ├── 00004.png
       ├──custom_data
            ├── img0.png
            ├── img1.png
            ├── img2.png
            ├── img3.png
            ├── img4.png
            ...

train model

python tools/main.py -c configs/stylegan_v2_256_ffhq.yaml

Inference

When you finish training, you need to use tools/extract_weight.py to extract the corresponding weights.

python tools/extract_weight.py output_dir/YOUR_TRAINED_WEIGHT.pdparams --net-name gen_ema --output YOUR_WEIGHT_PATH.pdparams

Then use applications/tools/styleganv2.py to get results

python tools/styleganv2.py --output_path stylegan01 --weight_path YOUR_WEIGHT_PATH.pdparams --size 256

Note: --size should be same with your config file.

Results

Random Samples:

Samples

Random Style Mixing:

Random Style Mixing

Reference

@inproceedings{Karras2019stylegan2,
  title     = {Analyzing and Improving the Image Quality of {StyleGAN}},
  author    = {Tero Karras and Samuli Laine and Miika Aittala and Janne Hellsten and Jaakko Lehtinen and Timo Aila},
  booktitle = {Proc. CVPR},
  year      = {2020}
}