RUB-SysSec
About DCT spectrum
Firstly，in the 'README.md' file, you note FFHQ is distributed in a cropped version, how to crop FFHQ image？Can you give me a code to process it. Secondly，when i use ‘compute_statistics.py’ file to generate DCT spectrum of a data set sampled from StyleGAN trained on FFHQ，i can not spot visible artifacts for the generated images. I don't know if there is something wrong with me，can you give me some advice？
"I can not spot visible artifacts for the generated images", me too. Please advise.
I download the pytorch version of stylegan and use Nvidia FFHQ pretrained model to generate 10,000 images. After running the 'compute_statistics.py', I observed a very smooth spectrum without any artifacts.
I download the pytorch version of stylegan and use Nvidia FFHQ pretrained model to generate 10,000 images. After running the 'compute_statistics.py', I observed a very smooth spectrum without any artifacts.
The original plots where made using the tensorflow version of StyleGAN I1). PyTorch might implement the upsampling techniques differently (see https://openaccess.thecvf.com/content/CVPR2022/html/Parmar_On_Aliased_Resizing_and_Surprising_Subtleties_in_GAN_Evaluation_CVPR_2022_paper.html).
BTW just because humans cannot see differences anymore, this does not mean that classifiers do not find it useful. Several newer works still use similar techniques (https://ojs.aaai.org/index.php/AAAI/article/view/19954).
Poor generalization, waste of time
The paper is three years old at this stage. Obviously newer models have adapted. See for example StyleGAN3 which specifically addressed upsampling problems..