AnyLiDAR: Any Sensor in One Model for LiDAR Generation

Existing LiDAR generators are typically trained on individual datasets, and directly pooling heterogeneous observations can degrade generation quality. We introduce AnyLiDAR, a unified generative model with geometry-conditioned ring modulation for multiple sensor domains. Our design exploits the circular scan order shared by full-sweep LiDAR observations while accounting for differences in ray layouts and depth and reflectance statistics. Dataset identity specifies the target distribution, while ray directions inform spatial attention and ring-context computation. The ring path aggregates current features and ray geometry along the scan order to predict position- and channel-dependent scale and shift coefficients for attention updates. This allows shared ring context to regulate the contribution of attention features according to measurement geometry and the current scan state. Trained on seven autonomous-driving datasets, AnyLiDAR achieves lower Fr\'echet Range Distance than all compared jointly trained baselines on five target domains, with a $10.2\%$ reduction over R2Flow on KITTI-360. Generative pre-training further improves SemanticKITTI segmentation by $9.3$ mIoU points over the same backbone trained from scratch.