PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Xiaobin Huang, Zilong Huang, Yang Luo, Hongchao Fan, Yiping Chen, Ting Han

arXiv:2608.05879v2cs.CV

TL;DR

Text-driven 3D generation typically treats urban exteriors and indoor scenes independently, leaving the correspondence needed for a coherent urban world unresolved. HoloWorld uses a continuously updated cross-scale context to generate exteriors and building-grounded interiors together, achieving stronger urban quality and indoor-outdoor consistency, including a 7.68% average AQS improvement over the SOTA baseline. Its current scope is limited to single-floor interiors without explicit vertical building structures.

  • Problem

    Existing urban and indoor generation methods remain separated and generally lack explicit correspondence between generated buildings and their interior realizations.

  • Method

    HoloWorld progressively updates a cross-scale world context from city planning to individual buildings, grounding indoor synthesis in exterior instances, appearances, assets, and footprints.

  • Results

    7.68%: HoloWorld improves average AQS over the SOTA baseline, achieves the highest average RDR score, and maintains strong building-level correspondence and cross-block continuity.

  • Takeaways & Limitations

    HoloWorld demonstrates that indoor and outdoor generation can form corresponding, coherent components of a unified 3D urban world.

  • Takeaways & Limitations

    The current framework focuses on single-floor interiors and does not explicitly model vertical building structures.

Abstract

from arXiv · show

Text-driven 3D generation has advanced rapidly in creating large-scale outdoor environments and detailed indoor scenes, but these domains are usually synthesized independently, lacking the correspondence required for a coherent urban world. We present HoloWorld, a unified indoor-outdoor urban world generation framework built on a continuously updated cross-scale world context. Initializing from a user description, HoloWorld progressively represents and updates the diverse world information, from city-scale planning to individual buildings, allowing generated interiors to maintain explicit correspondence with their associated exterior buildings. Conditioned on the evolving context and previously generated neighboring blocks, HoloWorld autoregressively generates urban exteriors with consistent spatial organization and visual identity across blocks. The generated exterior representations are further grounded in 3D building instances and footprints, enabling building-specific indoor generation with geometry-constrained layouts and inherited appearance characteristics. To our knowledge, HoloWorld is the first framework to unify indoor and outdoor generation within a coherent 3D urban world. Extensive experiments demonstrate that HoloWorld achieves superior urban exterior generation performance, improving the average AQS score over the SOTA by 7.68\% and obtaining the highest average RDR score, while maintaining strong building-level indoor-outdoor correspondence and cross-block continuity within a unified 3D urban world. Our project page: https://huangxb326.github.io/HoloWorld/.

Introduction

HoloWorld addresses the separation between urban exterior and indoor scene generation by treating them as corresponding parts of one coherent 3D urban world. Its cross-scale context propagates validated semantic, visual, spatial, and geometric information from city planning to individual buildings.

  • Existing methods generate plausible urban exteriors and indoor scenes, but generally lack explicit correspondence between buildings and their interiors.
  • A unified formulation requires each interior to inherit its building’s functional semantics and visual identity while respecting the exterior footprint.
  • The central challenge is preserving, propagating, and localizing contextual information across city, block, and building scales.
  • HoloWorld progressively refines a shared world context from a user description through city, block, and building generation.
  • 7.68%: HoloWorld improves average AQS over the SOTA baseline while achieving the highest average RDR score across diverse urban scenarios.
  • HoloWorld combines cross-scale context, coherent autoregressive block generation, and building-grounded indoor synthesis within one urban world.

Related Work

Related work has advanced urban-scale exterior generation and language-guided indoor synthesis, but these research directions remain largely separate. Existing systems address large environments, room and building interiors, or agentic scene construction without establishing unified indoor-outdoor correspondence.

  • Urban generation methods model large environments through compositional city representations, language-guided planning, controllable asset assembly, or tile-by-tile expansion.
  • Indoor generation has progressed from furniture arrangement to language-driven multi-room and building-scale synthesis with agentic planning and iterative refinement.

Problem Formulation

HoloWorld formulates unified indoor-outdoor generation as cross-scale 3D world generation, where spatially organized exteriors and corresponding interiors maintain semantic and geometric consistency. Its representation uses blocks, stable building instances, and building-specific interiors linked through a shared context.

  • Overview: Figure 2 organizes validated world information at world, block, and building levels before localizing it for building-specific indoor synthesis.
  • Objective: Given a text prompt, the objective is a coherent 3D urban world whose exteriors and corresponding interiors maintain semantic and geometric relationships.
  • Representation: The exterior is represented as an R×C grid of spatial blocks containing stable building instances, with a selected subset targeted for indoor generation.
  • Context: The cross-scale context bridges textual semantics, visual realization, geometric grounding, and indoor-outdoor generation while progressively refining the world from global planning to buildings.
  • Consistency: At the urban scale, context preserves continuity between neighboring regions; at the building scale, it links each interior to its exterior instance and footprint.

Cross-Scale World Context Representation

HoloWorld represents generation with a hierarchical living context rather than a simple output memory. The context extracts task-specific conditions, integrates only validated evidence, and connects world, block, and building information bidirectionally across generation.

  • The world context bridges semantic descriptions, visual appearances, spatial layouts, and building geometries throughout generation.
  • The context is decomposed into hierarchical world, block, and building levels, with building instances grounded at each generation stage.
  • World-level context defines city identity, block-level context preserves local spatial relationships, and building-level context provides conditions for corresponding indoor synthesis.
  • Information is inherited from higher levels to lower levels, while newly generated and validated evidence propagates back to enrich the context.
  • This bidirectional, continuously updated representation maintains global identity, inter-block continuity, and building-level exterior-interior correspondence.
  • At each generation step, task-specific conditions are projected from the current context, a generator produces results and evidence, and only validated results are integrated.

Exterior Realization and Building-Level Context Localization

HoloWorld plans city themes and inter-block relationships, generates exteriors autoregressively with neighboring-block context, and localizes that context to building instances and footprints for indoor generation.

  • The City Planning Module defines city themes, functional organization, inter-block relationships, and shared exterior-interior style references in Cworld.
  • Block-specific spatial layouts are generated from Cworld and each block plan, then stored for exterior generation.
  • Exterior blocks are generated autoregressively using world context, block design, and previously generated neighboring blocks to preserve boundary continuity.
  • Building regions in block images are associated with corresponding 3D instances, from which building identities and footprints are obtained.
  • The resulting building-level context integrates inherited semantics, appearance, instance identity, and geometric constraints for building-specific indoor synthesis.

Building-Level Context-Driven Indoor Generation

Building-level context conditions indoor synthesis with inherited world information, building-specific appearance and identity, and geometry, producing interiors aligned with their exterior structures.

  • Building-level context Cb combines inherited world information with building-specific appearance, identity, and geometry for indoor synthesis.
  • Indoor generation first constructs a structured interior program and uses exterior appearance and inherited style to form a building-specific asset library.
  • Hierarchical synthesis determines room layouts before progressively placing furniture, architectural elements, and smaller objects according to spatial dependencies.
  • The indoor formulation extracts programs, style conditions, assets, and footprint constraints from Cb, restricting interiors to the exterior building boundary.It anchors indoor synthesis to the corresponding exterior while preserving inherited urban context.

Experiments

Experiments evaluate HoloWorld’s urban exteriors, indoor-outdoor coherence, and cross-block continuity against established exterior and independent indoor-outdoor baselines using automated and human assessments.

  • HoloWorld is evaluated on generated cities spanning urban functions, architectural styles, and environmental themes, with 3 × 3 block grids used for fair comparison.
  • Urban exterior generation is compared with CityCraft, SynCity, and MajutsuCity using identical urban descriptions.
  • TRELLIS provides an independent-generation baseline that creates matched exterior and interior scenes separately from functional and stylistic descriptions.
  • Evaluation covers urban exterior quality, indoor-outdoor coherence, and cross-block continuity.
  • Exterior quality uses AQS and RDR across structural-view consistency, scene richness-complexity, material-texture fidelity, and lighting-atmosphere dimensions.Assessments use GPT-5.5-based judgments and human evaluations with 20 experts.
  • Indoor-outdoor coherence measures functional, visual, and spatial consistency, while cross-block evaluation examines roads, ground surfaces, building appearance, and rendering style across adjacent blocks.

Quantitative Comparison

HoloWorld improves urban exterior quality and coherence while maintaining building-level indoor-outdoor correspondence and cross-block continuity. Qualitative and quantitative comparisons show richer environments, stronger organization, and consistent visual identity.

  • HoloWorld achieves the best performance across all four evaluation dimensions under both AQS and RDR protocols.It also obtains the highest RDR scores in all four dimensions.
  • 9.38%, 4.65%, 9.00%, and 8.11% are HoloWorld’s relative AQS gains over MajutsuCity on SVC, SRC, MTF, and LA, respectively.
  • HoloWorld produces richer architectural, landscape, and public-space details while preserving coherent spatial organization and visual style across the city.It jointly synthesizes buildings, roads, parks, plazas, and layered vegetation with distinctive regional characteristics.
  • Compared with SynCity’s tile-based strategy, HoloWorld maintains more coherent organization and visual consistency across neighboring blocks while preserving detailed local structures.
  • HoloWorld’s interiors remain functionally, visually, and spatially consistent with corresponding exterior buildings through building-level correspondence.Interior organization follows footprints, functional layouts match building roles, and assets and materials inherit surrounding visual identity.

Ablation Studies

Ablations show that dynamic world-context updates are necessary for building-level indoor-outdoor correspondence, while autoregressive neighborhood conditioning supports cross-block continuity. Removing either component degrades the corresponding evaluation outcomes.

  • Dynamic World Context: The static configuration receives only the initial city description and uses a default rectangular boundary instead of localized, continuously updated building context.
  • Dynamic World Context: Freezing world-context updates reduces functional, visual, and spatial AQS from 7.67, 8.17, and 7.75 to 5.75, 4.58, and 3.25, respectively.Shape IoU also falls from 0.994 to 0.670 under GPT-5.5 evaluation.
  • Dynamic World Context: Without context updates, average human-evaluation AQS decreases from 8.21 to 5.54.The initial city description alone does not provide sufficient building-specific semantic, visual, and geometric constraints.
  • Autoregressive Neighborhood Conditioning: Removing autoregressive neighborhood conditioning lowers GPT-5.5 continuity AQS from 8.25 to 7.25 and RDR from 24.04 to 16.66.
  • Autoregressive Neighborhood Conditioning: Human evaluation likewise shows continuity AQS decreasing from 7.75 to 5.38 and RDR from 21.97 to 19.07 without neighboring-block context.Previously generated neighboring blocks provide contextual references for preserving local spatial relations and visual characteristics.

Conclusion

HoloWorld unifies indoor and outdoor urban scene generation through a continuously updated cross-scale world context, preserving building-level correspondence and continuity across neighboring blocks. Experiments report superior exterior quality and strong indoor-outdoor correspondence, while the current framework remains limited to single-floor interiors.

  • Conclusion: HoloWorld treats exteriors and interiors as corresponding realizations of one 3D world through a continuously updated cross-scale context.The context transfers semantic, visual, and spatial information from city-level planning to individual buildings.
  • Conclusion: Context-aware autoregressive generation preserves spatial and visual continuity across neighboring urban blocks.
  • Conclusion: Experiments show superior urban exterior quality and strong building-level correspondence between generated interiors and associated buildings.Ablations also verify the importance of the living context and autoregressive neighborhood conditioning.
  • Conclusion: The current framework focuses on single-floor interiors and does not explicitly model vertical building structures.Future work targets multi-floor architectural reasoning, richer structural constraints, and interactive physical validation.
Loading 2608.05879v2…