This paper proposes a novel wavelet-domain loss-based approach for training generative image super-resolution (SR) models to better control artifacts. The key idea is to train GAN-based SR models using wavelet-domain losses, which are more effective than RGB-domain or Fourier-space losses in distinguishing genuine image details from artifacts. The proposed method, called Wavelet-Guided Super-Resolution (WGSR), trains the discriminator only on high-frequency (HF) wavelet subbands and the generator using a fidelity loss over wavelet subbands to make it sensitive to the scale and orientation of structures. This approach enables the model to learn image details with different scales and orientations, leading to a better perception-distortion (PD) trade-off. The WGSR model outperforms existing state-of-the-art (SOTA) methods in both quantitative and qualitative evaluations. It achieves higher PSNR and better perceptual quality scores, as demonstrated by experiments on benchmark datasets such as Set5, Set14, BSD100, Urban100, and DIV2K. The model also shows superior performance in visual quality, as it can reconstruct genuine image details without hallucinations or artifacts. The results indicate that wavelet-domain losses are a suitable optimization objective for training GAN-SR models to achieve photo-realistic, high-quality, and accurate SR images. The paper also discusses the effectiveness of different wavelet families and the choice of decomposition levels in the wavelet domain. It shows that the Symlet "sym7" filter provides the best trade-off between objective and perceptual quality. Additionally, an ablation study is conducted to evaluate the impact of different loss terms, including fidelity, adversarial, and perceptual losses, on the performance of the WGSR model. The results demonstrate that using wavelet-domain losses leads to a better PD trade-off compared to traditional RGB-domain losses. The proposed method is generic and can be applied to any GAN-SR model. It provides a new approach to training SR models that can better control artifacts and achieve a better PD trade-off. The WGSR model is the first GAN-based SR model guided by wavelet-domain losses, and it shows significant improvements in both quantitative and qualitative performance compared to existing methods.This paper proposes a novel wavelet-domain loss-based approach for training generative image super-resolution (SR) models to better control artifacts. The key idea is to train GAN-based SR models using wavelet-domain losses, which are more effective than RGB-domain or Fourier-space losses in distinguishing genuine image details from artifacts. The proposed method, called Wavelet-Guided Super-Resolution (WGSR), trains the discriminator only on high-frequency (HF) wavelet subbands and the generator using a fidelity loss over wavelet subbands to make it sensitive to the scale and orientation of structures. This approach enables the model to learn image details with different scales and orientations, leading to a better perception-distortion (PD) trade-off. The WGSR model outperforms existing state-of-the-art (SOTA) methods in both quantitative and qualitative evaluations. It achieves higher PSNR and better perceptual quality scores, as demonstrated by experiments on benchmark datasets such as Set5, Set14, BSD100, Urban100, and DIV2K. The model also shows superior performance in visual quality, as it can reconstruct genuine image details without hallucinations or artifacts. The results indicate that wavelet-domain losses are a suitable optimization objective for training GAN-SR models to achieve photo-realistic, high-quality, and accurate SR images. The paper also discusses the effectiveness of different wavelet families and the choice of decomposition levels in the wavelet domain. It shows that the Symlet "sym7" filter provides the best trade-off between objective and perceptual quality. Additionally, an ablation study is conducted to evaluate the impact of different loss terms, including fidelity, adversarial, and perceptual losses, on the performance of the WGSR model. The results demonstrate that using wavelet-domain losses leads to a better PD trade-off compared to traditional RGB-domain losses. The proposed method is generic and can be applied to any GAN-SR model. It provides a new approach to training SR models that can better control artifacts and achieve a better PD trade-off. The WGSR model is the first GAN-based SR model guided by wavelet-domain losses, and it shows significant improvements in both quantitative and qualitative performance compared to existing methods.