Cutting-edge techniques in image-to-image translation are highly effective at learning the mappings between source and target domains using unpaired image datasets. However, despite their promising results, these methods often produce visual artifacts and struggle to capture high-level semantics, focusing primarily on low-level details. This issue arises because the generators typically fail to identify and utilize the most discriminative regions between domains, resulting in lower-quality generated images. In this paper, a novel approach:DMAG-GAN: Discriminator-derived Multimask Attention-Guided GAN is proposed. DMAG-GAN utilizes key attention information from the discriminator to direct the generator’s focus toward the most significant regions. Moreover, we implement multi-layer attention within the generator itself. This dual attention mechanism improves the GAN’s capacity to concentrate on essential semantic features, resulting in more realistic and higher-quality output images. We apply this method to the task of translating aerial satellite images to map images and record results based on this application. The experiments conducted indicate that the proposed approach outperforms conventional multimask AttentionGANs that rely solely on generator-derived attention mechanisms, demonstrating its effectiveness in producing higher-quality translations.