To See a World in a Living Context: Unified Indoor-Outdoor Urban World Generation

To See a World in a Living Context. Unified Indoor-Outdoor Urban World Generation. By Xiaobin Huang and five others, from Sun Yat-sen University and the Norwegian University of Science and Technology. Published on the twenty-ninth of August. No code repository is linked to it.

Text-driven 3D generation has become good at two things separately. Large outdoor environments on one side, detailed indoor scenes on the other. Each domain is usually synthesized on its own, and that independence is the difficulty. A city generated one way and its interiors generated another way have no correspondence between them. Step through a door and nothing connects the room you enter to the building you saw from the street. What is missing is not fidelity in either half. It is coherence across the scales that separate them.

HoloWorld proposes a single framework that generates indoors and outdoors together, and the thing holding them together is a cross-scale world context that is continuously updated. Rather than fixing a plan once and rendering from it, the context behaves as a living record. It initializes from a user description, then progressively represents and updates the world information as generation proceeds, carrying it from city-scale planning down to an individual building. Every later decision is conditioned on what that record already holds.

Generation then runs in two grounded stages. Exteriors come first: conditioned on the evolving context and on neighbouring blocks already generated, HoloWorld autoregressively produces urban exteriors that keep consistent spatial organisation and visual identity from block to block. Those exterior representations are then grounded in 3D building instances and footprints. That grounding is what makes the indoor stage building-specific. Layouts are geometry-constrained by the building they sit inside, and appearance characteristics are inherited from the exterior rather than invented, so an interior keeps explicit correspondence with the building it belongs to.

Extensive experiments report superior urban exterior generation. The average AQS score improves over the previous state of the art by seven point six eight percent, and the average RDR score is the highest obtained. Reported alongside those are strong building-level indoor-outdoor correspondence and cross-block continuity, held within one unified 3D urban world. The claim under test is not only that each half looks good on its own, but that the two halves agree with each other.

To the authors' knowledge this is the first framework to unify indoor and outdoor generation inside a coherent 3D urban world. The interesting part is where the coherence comes from. Not a larger model, and not an alignment step bolted on afterwards, but a context that persists and is updated while the world is being built, so the scales stay in correspondence by construction.

To See a World in a Living Context. Unified Indoor-Outdoor Urban World Generation. By Xiaobin Huang and five others, from Sun Yat-sen University and the Norwegian University of Science and Technology. Published on the twenty-ninth of August. This is the long version: the whole paper rather than the abstract.

The paper opens by quoting Blake: to see a World in a Grain of Sand. What it borrows from the line is a property of coherent environments, that a local observation should stay consistent with a larger global identity. Now consider what text-driven 3D generation actually produces today. You can generate a convincing city. You can generate a convincing room. Generate them separately and the room has nothing to do with the city. Walk up to a building with a jade roof and blocky voxel architecture, step inside, and the interior may belong to some other world entirely. Both halves were plausible. Neither was answerable to the other.

The two research directions have stayed apart for structural reasons rather than accidental ones. Urban generation methods build exteriors without modelling the internal spaces of individual buildings. Indoor generation methods produce isolated scenes without grounding them in any existing urban context. Even recent agent-based 3D systems rarely establish an explicit correspondence between a generated building and its interior realization. What the authors propose instead is a formulation. Every generated interior should correspond to a building instance, inherit that building's functional semantics and visual identity, and respect its footprint. That turns indoor-outdoor generation from independent scene synthesis into a world-consistent generation problem, and a world-consistent problem needs information to propagate across spatial scales.

The mechanism is a structure the authors call the living context. Formally, the cross-scale world context is decomposed into three hierarchical levels. World-level context encodes the identity of the city: urban semantics, functional organization, shared visual styles. Block-level context specializes that for one spatial region and maintains the local spatial relationships a coherent block needs. Building-level context localizes it further to a single building instance, integrating identity, appearance and geometric properties. Information is inherited downward, and validated evidence propagates back up. Two details make this a context rather than a memory. Each generation step extracts only what its task requires, through a projection, so no generator is handed the whole world. And results are gated: a step produces a candidate and new evidence, the context absorbs them only if a stage-specific validation passes, and otherwise the previous context is retained unchanged. Only reliable information moves forward.

Generation runs in two grounded stages. The exterior is produced autoregressively over a grid of blocks, each conditioned on world-level context, its own block-level design, and the neighbouring blocks already generated. Each block image becomes a 3D model, and those assemble over the grid into the urban exterior. Then comes the step that makes indoor generation possible at all. A building region inside a block image is associated with its 3D instance, which gives the building a stable identity and a recovered footprint. Exterior appearance is treated as design evidence rather than as observation of a hidden interior, which is the honest way to put it. That footprint then defines the spatial domain the interior must occupy, and hierarchical synthesis proceeds inside it, from room layout down through furniture and architectural elements to smaller objects.

The stated limitation is narrow, and precision about it matters. The current framework focuses on single-floor interiors, and does not explicitly model vertical building structures. For a system whose entire claim is that an interior corresponds to a particular building, that is a real boundary: a tower and a single-storey house present it with the same problem. Future work is aimed at multi-floor architectural reasoning, richer structural constraints, and interactive physical validation.

What to take from this is the shape of the fix rather than the size of the scores. The failure being addressed was never that either generator was weak. Both halves were already producing plausible output on their own. The failure was that nothing carried what one half had decided into the other half's inputs. The remedy is a structure that persists across the whole generation, that is written to only when a result validates, and that is projected down to exactly what each step needs. Coherence here is not a loss term, and not an alignment step bolted on at the end. It is a consequence of where the information was allowed to flow.