The paper introduces SeD, a semantic-aware discriminator for image super-resolution (SR) that enhances the generation of realistic and semantically consistent textures. Traditional discriminators in GAN-based SR methods are too coarse-grained, leading to unrealistic textures. SeD addresses this by incorporating semantic information from pretrained vision models (PVMs) into the discriminator, enabling it to learn fine-grained semantic-aware distributions. This is achieved through a semantic-aware fusion block (SeFB) that uses cross-attention to integrate semantic features into the discriminator. The semantic features are extracted from PVMs such as CLIP, and the SeFB warps these features to guide the discriminator in distinguishing real from fake images based on semantics. This approach improves the SR network's ability to generate photo-realistic and visually pleasing images. The method is integrated into both patch-wise and pixel-wise discriminators, and extensive experiments on classical and real-world SR tasks demonstrate its effectiveness. The results show that SeD outperforms existing methods in terms of perceptual quality metrics like LPIPS and PSNR, and it is compatible with various GAN-based SR frameworks such as ESRGAN, RealESRGAN, and BSRGAN. The proposed SeD provides a general and flexible solution for enhancing the semantic awareness in SR tasks.The paper introduces SeD, a semantic-aware discriminator for image super-resolution (SR) that enhances the generation of realistic and semantically consistent textures. Traditional discriminators in GAN-based SR methods are too coarse-grained, leading to unrealistic textures. SeD addresses this by incorporating semantic information from pretrained vision models (PVMs) into the discriminator, enabling it to learn fine-grained semantic-aware distributions. This is achieved through a semantic-aware fusion block (SeFB) that uses cross-attention to integrate semantic features into the discriminator. The semantic features are extracted from PVMs such as CLIP, and the SeFB warps these features to guide the discriminator in distinguishing real from fake images based on semantics. This approach improves the SR network's ability to generate photo-realistic and visually pleasing images. The method is integrated into both patch-wise and pixel-wise discriminators, and extensive experiments on classical and real-world SR tasks demonstrate its effectiveness. The results show that SeD outperforms existing methods in terms of perceptual quality metrics like LPIPS and PSNR, and it is compatible with various GAN-based SR frameworks such as ESRGAN, RealESRGAN, and BSRGAN. The proposed SeD provides a general and flexible solution for enhancing the semantic awareness in SR tasks.