This paper presents a novel approach for unsupervised pixel-level domain adaptation using generative adversarial networks (GANs). The proposed method, called PixelDA, learns a transformation in pixel space from one domain to another without requiring corresponding pairs from the two domains. The GAN-based model adapts source-domain images to appear as if they were drawn from the target domain. The method not only produces plausible samples but also outperforms state-of-the-art approaches on multiple unsupervised domain adaptation scenarios. The adaptation process generalizes to object classes unseen during training. The model is trained to adapt images from the source domain to look like they were sampled from the target domain while maintaining their original content. It uses a generator function that maps source domain images and noise vectors to adapted images. A discriminator function is used to distinguish between real and fake images, and a task-specific classifier assigns task-specific labels to images. The model is optimized using a minimax objective that combines domain loss, task loss, and content-similarity loss. The method is evaluated on several datasets, including MNIST, MNIST-M, USPS, and a variation of the LineMod dataset. The results show that PixelDA outperforms previous work on these datasets, particularly in the challenging "Synthetic Cropped LineMod to Cropped LineMod" scenario, where it reduces the mean angle error for pose estimation by more than half compared to the previous best result. The model is also able to generalize to unseen object classes and performs well in a semi-supervised setting with a small number of labeled target examples. The model's ability to separate style from content and its stability in training are key advantages. The approach is effective in adapting images from the source domain to the target domain, making it a promising method for unsupervised domain adaptation.This paper presents a novel approach for unsupervised pixel-level domain adaptation using generative adversarial networks (GANs). The proposed method, called PixelDA, learns a transformation in pixel space from one domain to another without requiring corresponding pairs from the two domains. The GAN-based model adapts source-domain images to appear as if they were drawn from the target domain. The method not only produces plausible samples but also outperforms state-of-the-art approaches on multiple unsupervised domain adaptation scenarios. The adaptation process generalizes to object classes unseen during training. The model is trained to adapt images from the source domain to look like they were sampled from the target domain while maintaining their original content. It uses a generator function that maps source domain images and noise vectors to adapted images. A discriminator function is used to distinguish between real and fake images, and a task-specific classifier assigns task-specific labels to images. The model is optimized using a minimax objective that combines domain loss, task loss, and content-similarity loss. The method is evaluated on several datasets, including MNIST, MNIST-M, USPS, and a variation of the LineMod dataset. The results show that PixelDA outperforms previous work on these datasets, particularly in the challenging "Synthetic Cropped LineMod to Cropped LineMod" scenario, where it reduces the mean angle error for pose estimation by more than half compared to the previous best result. The model is also able to generalize to unseen object classes and performs well in a semi-supervised setting with a small number of labeled target examples. The model's ability to separate style from content and its stability in training are key advantages. The approach is effective in adapting images from the source domain to the target domain, making it a promising method for unsupervised domain adaptation.