This paper introduces the Energy-Based Generative Adversarial Network (EBGAN), a novel approach to generative adversarial networks (GANs) that views the discriminator as an energy function. In EBGAN, the discriminator is trained to assign low energy values to data points near the data manifold and higher energy values to other regions. The generator, in turn, is trained to produce samples with low energy, which are then used to train the discriminator. This framework allows for a wide variety of architectures and loss functions beyond the traditional binary classifier used in GANs. The EBGAN model is based on the concept of energy-based models, where the goal is to shape the energy surface such that desired configurations have low energy and incorrect ones have high energy. The generator is trained to produce samples that minimize the energy, while the discriminator is trained to maximize the energy of these samples. This approach leads to more stable training compared to traditional GANs and allows for the use of auto-encoder architectures, where the energy is the reconstruction error. The paper presents a theoretical analysis of the EBGAN model, showing that under a simple hinge loss, when the system reaches convergence, the generator produces samples that follow the underlying data distribution. The authors also demonstrate that a single-scale architecture can be trained to generate high-resolution images. Experiments on various datasets, including MNIST, LSUN, CelebA, and ImageNet, show that EBGANs outperform traditional GANs in terms of training stability and image generation quality. The EBGAN framework is also shown to be effective in semi-supervised learning tasks, where it improves the performance of the Ladder Network by providing additional contrastive samples. The paper concludes that EBGANs offer a more flexible and stable approach to generative modeling, with the potential for further improvements in conditional settings and broader applications.This paper introduces the Energy-Based Generative Adversarial Network (EBGAN), a novel approach to generative adversarial networks (GANs) that views the discriminator as an energy function. In EBGAN, the discriminator is trained to assign low energy values to data points near the data manifold and higher energy values to other regions. The generator, in turn, is trained to produce samples with low energy, which are then used to train the discriminator. This framework allows for a wide variety of architectures and loss functions beyond the traditional binary classifier used in GANs. The EBGAN model is based on the concept of energy-based models, where the goal is to shape the energy surface such that desired configurations have low energy and incorrect ones have high energy. The generator is trained to produce samples that minimize the energy, while the discriminator is trained to maximize the energy of these samples. This approach leads to more stable training compared to traditional GANs and allows for the use of auto-encoder architectures, where the energy is the reconstruction error. The paper presents a theoretical analysis of the EBGAN model, showing that under a simple hinge loss, when the system reaches convergence, the generator produces samples that follow the underlying data distribution. The authors also demonstrate that a single-scale architecture can be trained to generate high-resolution images. Experiments on various datasets, including MNIST, LSUN, CelebA, and ImageNet, show that EBGANs outperform traditional GANs in terms of training stability and image generation quality. The EBGAN framework is also shown to be effective in semi-supervised learning tasks, where it improves the performance of the Ladder Network by providing additional contrastive samples. The paper concludes that EBGANs offer a more flexible and stable approach to generative modeling, with the potential for further improvements in conditional settings and broader applications.