GenAD is a generative end-to-end autonomous driving framework that directly produces planning results from raw sensors. The framework models autonomous driving as a trajectory generation problem, enabling comprehensive and efficient scene descriptions through instance-centric scene representations. It incorporates high-order interactions between the ego vehicle and other agents, and integrates semantic map information to enhance motion prediction and planning. The framework uses a variational autoencoder to model the trajectory prior in a latent space, and a gated recurrent unit to generate future trajectories. The proposed GenAD achieves state-of-the-art performance on the nuScenes benchmark, demonstrating high efficiency and accuracy in vision-centric end-to-end autonomous driving. The framework also incorporates auxiliary tasks such as map segmentation and 3D object detection to improve overall performance. Extensive experiments show that GenAD outperforms existing methods in motion prediction, planning, and perception tasks, with a strong performance/speed trade-off. The framework is trained end-to-end and can generate future trajectories in a structured latent space, considering the prior of realistic trajectories to produce high-quality predictions. The results show that GenAD produces better and safer trajectories than existing methods in various scenarios, including complex traffic situations. The framework is efficient and effective, demonstrating the potential of generative modeling in end-to-end autonomous driving.GenAD is a generative end-to-end autonomous driving framework that directly produces planning results from raw sensors. The framework models autonomous driving as a trajectory generation problem, enabling comprehensive and efficient scene descriptions through instance-centric scene representations. It incorporates high-order interactions between the ego vehicle and other agents, and integrates semantic map information to enhance motion prediction and planning. The framework uses a variational autoencoder to model the trajectory prior in a latent space, and a gated recurrent unit to generate future trajectories. The proposed GenAD achieves state-of-the-art performance on the nuScenes benchmark, demonstrating high efficiency and accuracy in vision-centric end-to-end autonomous driving. The framework also incorporates auxiliary tasks such as map segmentation and 3D object detection to improve overall performance. Extensive experiments show that GenAD outperforms existing methods in motion prediction, planning, and perception tasks, with a strong performance/speed trade-off. The framework is trained end-to-end and can generate future trajectories in a structured latent space, considering the prior of realistic trajectories to produce high-quality predictions. The results show that GenAD produces better and safer trajectories than existing methods in various scenarios, including complex traffic situations. The framework is efficient and effective, demonstrating the potential of generative modeling in end-to-end autonomous driving.