Spatial reconstruction and generation is the productive counterpart to spatial perception and reasoning: rather than analyzing a scene that is given, it synthesizes spatial content, recovering or creating the geometry, appearance, and dynamics of objects, scenes, and worlds from conditions such as text, images, partial observations, or explicit constraints. Where perception (Chapter 6) maps observations into structured spatial representations and reasoning (Chapter 7) draws inferences over them, generation closes the loop by mapping intent back into concrete spatial artifacts, a capability central to embodied agents, content creation, simulation, and world modeling.
The chapter is organized along an axis of increasing output structure and dimensionality, from flat images to volumetric objects and scenes, to video, embodied motion, and full four-dimensional space-time content, and finally to world models and to programmatic generation that emits interpretable, executable structure. What holds this range together is a single recurring backbone: an encoder maps conditions into a latent, a generator samples structure, and a decoder or renderer produces the final artifact, specialized to each output domain. Read this way, the subfields are less seven separate problems than one conditional-generation abstraction instantiated against progressively richer codomains.