SceneGenesis: 3D Scene Synthesis via Semantic Structural Priors and Mesh-Guided Video-Geometry Fusion

Generating high-quality, controllable, and structurally consistent 3D scenes is a fundamental yet challenging task, especially in complex multi-object environments. We present \textbf{SceneGenesis}, a unified framework for 3D scene synthesis that systematically integrates semantic structural priors with mesh-guided video-geometry fusion. The process begins with a \textbf{semantic structural initialization module}, which leverages large language models to convert textual scene prompts into category-aware object descriptions. These are transformed into structured meshes by combining procedural approximations for large-scale objects and pretrained mesh generators for fine-grained assets, enabling precise layout control and scene scalability. To synthesize rich and style-controllable appearances, we render depth and semantic maps from the initialized scene and condition a pretrained video diffusion model to generate multi-view video sequences with geometry-awareness, where a consistency-guided latent fusion strategy further enhances temporal consistency across long sequences. Crucially, we introduce a \textbf{mesh-guided video-geometry fusion module} that reconstructs coherent 3D Gaussian scenes by aligning mesh priors with video outputs. This module incorporates mesh-conditioned fragment initialization, progressive geometric refinement, and structure-aware optimization, significantly enhancing global geometric fidelity and visual realism. Extensive experiments demonstrate that \textbf{SceneGenesis} enables flexible style variation and object-level editing while achieving superior controllability, scalability, and 3D structural quality, offering an effective solution for 3D scene synthesis.