This paper introduces Generative Adversarial Nets (GANs), a new framework for estimating generative models through an adversarial process. The framework involves training two models: a generative model G that captures the data distribution, and a discriminative model D that estimates the probability that a sample came from the training data rather than G. The training process for G is to maximize the probability of D making a mistake, which corresponds to a minimax two-player game. In the space of arbitrary functions G and D, a unique solution exists, with G recovering the training data distribution and D equal to 1/2 everywhere. When G and D are defined by multilayer perceptrons, the entire system can be trained with backpropagation, without the need for Markov chains or unrolled approximate inference networks. Experiments demonstrate the potential of the framework through qualitative and quantitative evaluation of the generated samples. The adversarial nets framework is most straightforward when both models are multilayer perceptrons. The generator G is trained to produce samples that are indistinguishable from real data, while the discriminator D is trained to distinguish between real and generated samples. The two models compete, with G trying to fool D and D trying to detect fakes. This competition drives both models to improve until the generated samples are indistinguishable from real data. Theoretical analysis shows that the training criterion allows one to recover the data generating distribution as G and D are given enough capacity. The framework can be implemented using an iterative, numerical approach, with alternating steps of optimizing D and G. The training objective for D is to maximize the probability of assigning the correct label to both training examples and samples from G. The generator G is trained to minimize log(1-D(G(z))), but in practice, it is more effective to train G to maximize log D(G(z)). The framework has several advantages, including the ability to train without Markov chains, the use of backpropagation for gradient computation, and the ability to incorporate a wide variety of functions into the model. However, it also has disadvantages, such as the lack of an explicit representation of the generative distribution and the need for careful synchronization between G and D during training. The framework has been tested on various datasets, including MNIST, the Toronto Face Database, and CIFAR-10, with promising results. The results suggest that the adversarial framework has the potential to be a powerful tool for generative modeling.This paper introduces Generative Adversarial Nets (GANs), a new framework for estimating generative models through an adversarial process. The framework involves training two models: a generative model G that captures the data distribution, and a discriminative model D that estimates the probability that a sample came from the training data rather than G. The training process for G is to maximize the probability of D making a mistake, which corresponds to a minimax two-player game. In the space of arbitrary functions G and D, a unique solution exists, with G recovering the training data distribution and D equal to 1/2 everywhere. When G and D are defined by multilayer perceptrons, the entire system can be trained with backpropagation, without the need for Markov chains or unrolled approximate inference networks. Experiments demonstrate the potential of the framework through qualitative and quantitative evaluation of the generated samples. The adversarial nets framework is most straightforward when both models are multilayer perceptrons. The generator G is trained to produce samples that are indistinguishable from real data, while the discriminator D is trained to distinguish between real and generated samples. The two models compete, with G trying to fool D and D trying to detect fakes. This competition drives both models to improve until the generated samples are indistinguishable from real data. Theoretical analysis shows that the training criterion allows one to recover the data generating distribution as G and D are given enough capacity. The framework can be implemented using an iterative, numerical approach, with alternating steps of optimizing D and G. The training objective for D is to maximize the probability of assigning the correct label to both training examples and samples from G. The generator G is trained to minimize log(1-D(G(z))), but in practice, it is more effective to train G to maximize log D(G(z)). The framework has several advantages, including the ability to train without Markov chains, the use of backpropagation for gradient computation, and the ability to incorporate a wide variety of functions into the model. However, it also has disadvantages, such as the lack of an explicit representation of the generative distribution and the need for careful synchronization between G and D during training. The framework has been tested on various datasets, including MNIST, the Toronto Face Database, and CIFAR-10, with promising results. The results suggest that the adversarial framework has the potential to be a powerful tool for generative modeling.