AI Day直播 | WorldSplat：用于自动驾驶的高斯中心前馈4D场景生成

Core Viewpoint - The article discusses the advancements in driving scene generation and reconstruction technologies, highlighting the introduction of WorldSplat, a novel feed-forward 4D driving scene generation framework that effectively generates consistent multi-trajectory videos [3][8]. Summary by Sections Driving Scene Generation and Reconstruction - Recent progress in driving scene generation and reconstruction technologies shows significant potential in enhancing autonomous driving system performance by generating scalable and controllable training data [3]. - Existing generation methods primarily focus on synthesizing diverse and high-fidelity driving videos but struggle with 3D consistency and sparse viewpoint coverage, limiting their ability to support high-quality new viewpoint synthesis (NVS) [3]. Introduction of WorldSplat - WorldSplat is introduced as a solution to the challenges between scene generation and reconstruction, developed by research teams from Nankai University [3]. - The framework employs two key steps: (1) it integrates a multi-modal information fusion 4D perception latent diffusion model to generate pixel-aligned 4D Gaussian distributions in a feed-forward manner; (2) it utilizes an enhanced video diffusion model to optimize new viewpoint videos rendered from these Gaussian distributions [3]. Experimental Results - Extensive experiments conducted on benchmark datasets demonstrate that WorldSplat can effectively generate high-fidelity, spatiotemporally consistent multi-trajectory new viewpoint driving videos [3][8].