This paper introduces InfoGAN, an extension of Generative Adversarial Networks (GANs) that learns disentangled representations in an unsupervised manner. InfoGAN maximizes the mutual information between a small subset of latent variables and the observations, enabling it to discover meaningful and interpretable representations. The method is effective on various datasets, including MNIST, CelebA, and SVHN, where it successfully disentangles writing styles from digit shapes, pose from lighting in 3D images, and background digits from central digits. InfoGAN's approach is competitive with supervised methods and adds minimal computational overhead to GANs. The paper also discusses related work, reviews GANs, and provides a detailed derivation of the InfoGAN objective, including a variational lower bound for mutual information. Experimental results demonstrate InfoGAN's ability to learn high-quality, disentangled representations, making it a promising tool for unsupervised representation learning.This paper introduces InfoGAN, an extension of Generative Adversarial Networks (GANs) that learns disentangled representations in an unsupervised manner. InfoGAN maximizes the mutual information between a small subset of latent variables and the observations, enabling it to discover meaningful and interpretable representations. The method is effective on various datasets, including MNIST, CelebA, and SVHN, where it successfully disentangles writing styles from digit shapes, pose from lighting in 3D images, and background digits from central digits. InfoGAN's approach is competitive with supervised methods and adds minimal computational overhead to GANs. The paper also discusses related work, reviews GANs, and provides a detailed derivation of the InfoGAN objective, including a variational lower bound for mutual information. Experimental results demonstrate InfoGAN's ability to learn high-quality, disentangled representations, making it a promising tool for unsupervised representation learning.