CG-BEV: Conditional Generation of Bird's-Eye-View Segmentation Using BEV-optimised Diffusion

Bird's-Eye-View (BEV) representation is a widely adopted approach in autonomous driving for modelling complex spatial layouts. However, BEV in camera-based systems is constructed from perspective images, which inherently lose information during 3D-to-2D projection. This information loss is irreversible regardless of how powerful the BEV decoder is, and it is particularly severe in distant or occluded regions. Existing solutions either rely on costly external priors, such as HD maps or large pretrained world models, or tightly couple the diffusion module with a specific perception backbone, which limits their deployment flexibility and practical viability. We introduce CG-BEV, a Conditional Generation framework that enhances representative BEV perception families by training only a lightweight ControlNet-style enhancer on top of a BEV-optimised latent diffusion model. The control signal is constructed from BEV features and segmentation maps already produced by the frozen perception model, which makes CG-BEV easy to train and readily integrable into existing pipelines. Experiments on the nuScenes dataset validate the effectiveness of CG-BEV under multiple evaluation metrics. Our method achieves up to 5.47 miou point improvement across various BEV segmentation models and supports 512$\times$512 resolution semantic map generation with a manageable inference overhead, making it a practical add-on for existing BEV perception pipelines. Code and pretrained models will be released.