This paper presents a novel approach to training generative image super-resolution (SR) models using wavelet-domain losses, which enables better control over artifacts and improves the balance between fidelity and perceptual quality. The authors argue that existing methods often generate artifacts and hallucinations while reconstructing high-frequency (HF) image details. To address this, they propose training GAN-based SR models using wavelet-domain losses, which allow for better characterization of genuine HF details versus artifacts compared to RGB-domain or Fourier-space losses. Specifically, the discriminator is trained only on the HF wavelet sub-bands, and the generator is trained using a fidelity loss over these sub-bands to make it sensitive to the scale and orientation of structures. Extensive experimental results demonstrate that the proposed model, named WGSR, achieves a better perception-distortion trade-off according to multiple objective measures and visual evaluations. The paper also discusses the rationale for using wavelet-domain losses, the architecture of the proposed framework, and the experimental setup. The authors conclude that their method outperforms state-of-the-art methods in both quantitative and qualitative evaluations, providing a more realistic and high-quality SR image reconstruction.This paper presents a novel approach to training generative image super-resolution (SR) models using wavelet-domain losses, which enables better control over artifacts and improves the balance between fidelity and perceptual quality. The authors argue that existing methods often generate artifacts and hallucinations while reconstructing high-frequency (HF) image details. To address this, they propose training GAN-based SR models using wavelet-domain losses, which allow for better characterization of genuine HF details versus artifacts compared to RGB-domain or Fourier-space losses. Specifically, the discriminator is trained only on the HF wavelet sub-bands, and the generator is trained using a fidelity loss over these sub-bands to make it sensitive to the scale and orientation of structures. Extensive experimental results demonstrate that the proposed model, named WGSR, achieves a better perception-distortion trade-off according to multiple objective measures and visual evaluations. The paper also discusses the rationale for using wavelet-domain losses, the architecture of the proposed framework, and the experimental setup. The authors conclude that their method outperforms state-of-the-art methods in both quantitative and qualitative evaluations, providing a more realistic and high-quality SR image reconstruction.