This project investigates urban block reconstruction from cross-view satellite and street-level imagery, formulating the problem within a sparse-view image-based learning setting. Rather than relying on dense multi-view capture, the framework operates under realistic conditions of asymmetric and heterogeneous viewpoints, jointly reasoning over top-down satellite context and inside-out panoramic street imagery. The approach adapts large reconstruction models to multi-object, block-scale scenes and evaluates how performance-relevant urban geometry can be inferred from incomplete visual coverage. Supported by a coupled synthetic dataset for supervised training and evaluation, the work clarifies both the capabilities and limitations of recovering environmentally meaningful block-scale form from sparse-view urban imagery.