DecompDreamer: A Composition-Aware Curriculum for Structured 3D Asset Generation
DecompDreamer introduces a staged optimization strategy to solve gradient conflicts in compositional text-to-3D generation. It first builds a global structure by optimizing inter-object relationships, then refines individual components using targeted losses. This curriculum approach ensures stable convergence and produces higher-fidelity, geometrically consistent 3D scenes compared to prior methods.