Skip to content

Latest commit

 

History

History
110 lines (80 loc) · 14.9 KB

README.md

File metadata and controls

110 lines (80 loc) · 14.9 KB
Raw

StyleGANv2

Analyzing and Improving the Image Quality of Stylegan

Abstract

The style-based GAN architecture (StyleGAN) yields state-of-the-art results in data-driven unconditional generative image modeling. We expose and analyze several of its characteristic artifacts, and propose changes in both model architecture and training methods to address them. In particular, we redesign the generator normalization, revisit progressive growing, and regularize the generator to encourage good conditioning in the mapping from latent codes to images. In addition to improving image quality, this path length regularizer yields the additional benefit that the generator becomes significantly easier to invert. This makes it possible to reliably attribute a generated image to a particular network. We furthermore visualize how well the generator utilizes its output resolution, and identify a capacity problem, motivating us to train larger models for additional quality improvements. Overall, our improved model redefines the state of the art in unconditional image modeling, both in terms of existing distribution quality metrics as well as perceived image quality.

Results and Models

Results (compressed) from StyleGAN2 config-f trained by MMGeneration
Model Comment FID50k P&R50k Config Download
stylegan2_config-f_ffhq_1024 official weight 2.8134 62.856/49.400 config model
stylegan2_config-f_lsun-car_384x512 official weight 5.4316 65.986/48.190 config model
stylegan2_config-f_horse_256 official weight - - config model
stylegan2_config-f_church_256 official weight - - config model
stylegan2_config-f_cat_256 official weight - - config model
stylegan2_config-f_ffhq_256 our training 3.992 69.012/40.417 config model
stylegan2_config-f_ffhq_1024 our training 2.8185 68.236/49.583 config model
stylegan2_config-f_lsun-car_384x512 our training 2.4116 66.760/50.576 config model

FP16 Support and Experiments

Currently, we have supported FP16 training for StyleGAN2, and here are the results for the mixed-precision training. (Experiments for FFHQ1024 will come soon.)

Evaluation FID for FP32 and FP16 training

As shown in the figure, we provide 3 ways to do mixed-precision training for StyleGAN2:

  • stylegan2_c2_fp16_PL-no-scaler: In this setting, we try our best to follow the official FP16 implementation in StyleGAN2-ADA. Similar to the official version, we only adopt FP16 training for the higher-resolution feature maps (the last 4 stages in G and the first 4 stages). Note that we do not adopt the clamp way to avoid gradient overflow used in the official implementation. We use the autocast function from torch.cuda.amp package.
  • stylegan2_c2_fp16-globalG-partialD_PL-R1-no-scaler: In this config, we try to adopt mixed-precision training for the whole generator, but in partial discriminator (the first 4 higher-resolution stages). Note that we do not apply the loss scaler in the path length loss and gradient penalty loss. Because we always meet divergence after adopting the loss scaler to scale the gradient in these two losses.
  • stylegan2_c2_apex_fp16_PL-R1-no-scaler: In this setting, we adopt the APEX toolkit to implement mixed-precision training with multiple loss/gradient scalers. In APEX, you can assign different loss scalers for the generator and the discriminator respectively. Note that we still ignore the gradient scaler in the path length loss and gradient penalty loss.
Model Comment Dataset FID50k Config Download
stylegan2_config-f_ffhq_256 baseline FFHQ256 3.992 config ckpt
stylegan2_c2_fp16_partial-GD_PL-no-scaler_ffhq_256_b4x8_800k partial layers in fp16 FFHQ256 4.331 config ckpt
stylegan2_c2_fp16-globalG-partialD_PL-R1-no-scaler_ffhq_256_b4x8_800k the whole G in fp16 FFHQ256 4.362 config ckpt
stylegan2_c2_apex_fp16_PL-R1-no-scaler_ffhq_256_b4x8_800k the whole G&D in fp16 + two loss scaler FFHQ256 4.614 config ckpt

In addition, we also provide QuickTestImageDataset to users for quickly checking whether the code can be run correctly. It's more important for FP16 experiments, because some cuda operations may no support mixed precision training. Esepcially for APEX, you can use this config in your local machine by running:

bash tools/dist_train.sh \
  configs/styleganv2/stylegan2_c2_apex_fp16_quicktest_ffhq_256_b4x8_800k.py 1 \
  --work-dir ./work_dirs/quick-test

With a similar way, users can switch to config for partial-GD and config for globalG-partialD to test the other two mixed precision training configuration.

Note that to use the APEX toolkit, you have to installed it following the official guidance. (APEX is not included in our requirements.) If you are using GPUs without tensor core, you would better to switch to the newer PyTorch version (>= 1.7,0). Otherwise, the APEX installation or running may meet several bugs.

About Different Implementations of FID Metric

Model Comment FID50k FID Version Config Download
stylegan2_config-f_ffhq_1024 official weight 2.8732 Tero's StyleGAN config model | FID-Reals
stylegan2_config-f_ffhq_1024 our training 2.9413 Tero's StyleGAN config model | FID-Reals
stylegan2_config-f_ffhq_1024 official weight 2.8134 Our PyTorch config model | FID-Reals
stylegan2_config-f_ffhq_1024 our training 2.8185 Our PyTorch config model | FID-Reals

In this table, we observe that the FID with Tero's inception network is similar to that with PyTorch Inception (in MMGeneration). Thus, we use the FID with PyTorch's Inception net (but the weight is not the official model zoo) by default. Because it can be run on different PyTorch versions. If you use Tero's Inception net, your PyTorch must meet >=1.6.0.

More precalculated inception pickle files are listed here:

  • FFHQ 256x256 real inceptions, PyTorch InceptionV3. download
  • LSUN-Car 384x512 real inceptions, PyTorch InceptionV3. download

About Different Implementation and Setting of PR Metric

Model P&R Details Precision Recall
stylegan2_config-f_ffhq_1024 (official weight) use Tero's VGG16, P&R50k_full 67.876 49.299
stylegan2_config-f_ffhq_1024 (official weight) use Tero's VGG16, P&R50k 62.856 49.400
stylegan2_config-f_ffhq_1024 (official weight) use PyTorch's VGG16, P&R50k_full 67.662 55.460

As shown in this table, P&R50k_full is the metric used in StyleGANv1 and StyleGANv2. full indicates that we use the whole dataset for extracting the real distribution, e.g., 70000 images in FFHQ dataset. However, adopting the VGG16 provided from Tero requires that your PyTorch version must fulfill >=1.6.0. Be careful about using the PyTorch's VGG16 to extract features, which will cause higher precision and recall.

Citation

@inproceedings{karras2020analyzing,
  title={Analyzing and improving the image quality of stylegan},
  author={Karras, Tero and Laine, Samuli and Aittala, Miika and Hellsten, Janne and Lehtinen, Jaakko and Aila, Timo},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={8110--8119},
  year={2020},
  url={https://openaccess.thecvf.com/content_CVPR_2020/html/Karras_Analyzing_and_Improving_the_Image_Quality_of_StyleGAN_CVPR_2020_paper.html},
}