The introduction of high-quality image generation models, particularly the StyleGAN family, provides a powerful tool to synthesize and manipulate images. However, existing models are built upon high-quality (HQ) data as desired outputs, making them unfit for in-the-wild low-quality (LQ) images, which are common inputs for manipulation. In this work, we bridge this gap by proposing a novel GAN structure that allows for generating images with controllable quality. The network can synthesize various image degradation and restore the sharp image via a quality control code. Our proposed QC-StyleGAN can directly edit LQ images without altering their quality by applying GAN inversion and manipulation techniques. It also provides for free an image restoration solution that can handle various degradations, including noise, blur, compression artifacts, and their mixtures. Finally, we demonstrate numerous other applications such as image degradation synthesis, transfer, and interpolation.

Table 1: Hyperparameters used in each model training.Parameter FFHQ AFHQ-Cat LSUN-Church Resolution 256 256 512 512 256 256 Number of GPUs 8 8 8 Training length 5M 5M 5M Minibatch size 64 64 64 Minibatch stddev 8 8 8 Feature maps Resnet D Training time 1.5 days 3.5 days 1.5 days Finally, we consider the entire DegradBlock: DB (f, k q) = π (ϕ ( r Thus, the lemma in Equation 10 holds. Karras et al., from which we inherit most of the training details, including weight demodulation, path We also report other details for each training in Table 1. For the image restoration task, we train pSp and run PTI on a single Nvidia V100. First, our reconstructed images can easily be converted to their sharp version. Moreover, since QC-StyleGAN models both sharp and degraded images, the inversion results often stay in-distribution, allowing good editing results.

The introduction of high-quality image generation models, particularly the Style-GAN family, provides a powerful tool to synthesize and manipulate images.

The introduction of high-quality image generation models, particularly the StyleGAN family, provides a powerful tool to synthesize and manipulate images. However, existing models are built upon high-quality (HQ) data as desired outputs, making them unfit for in-the-wild low-quality (LQ) images, which are common inputs for manipulation. In this work, we bridge this gap by proposing a novel GAN structure that allows for generating images with controllable quality. The network can synthesize various image degradation and restore the sharp image via a quality control code. Our proposed QC-StyleGAN can directly edit LQ images without altering their quality by applying GAN inversion and manipulation techniques.

The introduction of high-quality image generation models, particularly the Style-GAN family, provides a powerful tool to synthesize and manipulate images. However, existing models are built upon high-quality (HQ) data as desired outputs, making them unfit for in-the-wild low-quality (LQ) images, which are common inputs for manipulation. In this work, we bridge this gap by proposing a novel GAN structure that allows for generating images with controllable quality. The network can synthesize various image degradation and restore the sharp image via a quality control code. Our proposed QC-StyleGAN can directly edit LQ images without altering their quality by applying GAN inversion and manipulation techniques. It also provides for free an image restoration solution that can handle various degradations, including noise, blur, compression artifacts, and their mixtures. Finally, we demonstrate numerous other applications such as image degradation synthesis, transfer, and interpolation.