OccGen is a generative model for 3D semantic occupancy prediction in autonomous driving. It adopts a "noise-to-occupancy" paradigm, progressively inferring and refining the occupancy map by predicting and eliminating noise from a random Gaussian distribution. The model consists of a conditional encoder that processes multi-modal inputs and a progressive refinement decoder that applies diffusion denoising using multi-modal features as conditions. OccGen can generate occupancy maps in a coarse-to-fine manner, producing more detailed predictions. It outperforms existing methods on several benchmarks, achieving significant improvements in mIoU metrics under different settings. OccGen also provides uncertainty estimates, a capability not available in discriminative models. The model is efficient, with comparable latency to single forward methods, and can be adapted to different trade-offs between speed and accuracy. Experiments show that OccGen achieves state-of-the-art performance on nuScenes-Occupancy and SemanticKITTI, demonstrating its effectiveness in semantic occupancy prediction.OccGen is a generative model for 3D semantic occupancy prediction in autonomous driving. It adopts a "noise-to-occupancy" paradigm, progressively inferring and refining the occupancy map by predicting and eliminating noise from a random Gaussian distribution. The model consists of a conditional encoder that processes multi-modal inputs and a progressive refinement decoder that applies diffusion denoising using multi-modal features as conditions. OccGen can generate occupancy maps in a coarse-to-fine manner, producing more detailed predictions. It outperforms existing methods on several benchmarks, achieving significant improvements in mIoU metrics under different settings. OccGen also provides uncertainty estimates, a capability not available in discriminative models. The model is efficient, with comparable latency to single forward methods, and can be adapted to different trade-offs between speed and accuracy. Experiments show that OccGen achieves state-of-the-art performance on nuScenes-Occupancy and SemanticKITTI, demonstrating its effectiveness in semantic occupancy prediction.