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

Xiaobin Huang1,*Zilong Huang1,*Yang Luo1Hongchao Fan2Yiping Chen1,†Ting Han1,†

1 Sun Yat-sen University

2 Norwegian University of Science and Technology

* Equal Contribution† Corresponding authors

Abstract

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.

HoloWorld compared with methods that generate indoor and outdoor scenes separately
Figure 1: Existing methods generate indoor and outdoor scenes separately. HoloWorld instead propagates a cross-scale world context from the urban world through its blocks to individual buildings, establishing explicit building–interior correspondence and preserving semantic, visual, and spatial coherence.

Method

World

Global identity, urban semantics, shared visual language.

Block

Local organization conditioned on surrounding city regions.

Building

A grounded instance with function, appearance, and footprint.

Interior

A corresponding space that belongs to its exterior building.

Overview of the HoloWorld generation pipeline
Figure 2: Overview of HoloWorld. (a) An evolving cross-scale world context organizes validated information at world, block, and building levels. (b) Context-driven outdoor generation and instance grounding localize this context to individual buildings, where it guides building-specific indoor synthesis under appearance, asset, and footprint constraints. (c) The resulting city, blocks, and interiors form explicitly corresponding parts of a unified 3D urban world.

Results

Qualitative comparison between HoloWorld and prior city generation methods
Figure 3: Qualitative comparison under identical urban descriptions, with two views per scene. HoloWorld generates richer architectural and landscape details while preserving more coherent spatial organization and visual style across the city.
HoloWorld living-context and autoregressive-neighborhood ablation results
Figure 4: Qualitative results and ablations of HoloWorld. (A) The full model generates an interior that remains functionally, visually, and spatially consistent with its corresponding exterior building, whereas a static world context weakens this correspondence. (B) Autoregressive neighborhood conditioning preserves cross-block spatial and visual continuity; removing it introduces visible boundary discontinuities highlighted by the red boxes.

Table 1: Quantitative comparison of urban exterior generation. All methods receive identical urban descriptions as inputs. GPT-5.5-based and human evaluations are reported separately. The best results are in bold.

Table 1: Quantitative comparison of urban exterior generation. All methods receive identical urban descriptions as inputs. GPT-5.5-based and human evaluations are reported separately. The best results are in bold.
MethodAQSRDR
SVC↑SRC↑MTF↑LA↑SVC↑SRC↑MTF↑LA↑
GPTHum.GPTHum.GPTHum.GPTHum.GPTHum.GPTHum.GPTHum.GPTHum.
CityCraft7.117.177.226.004.895.005.895.0023.6519.2120.7619.6119.9014.9917.3515.97
SynCity7.707.178.207.834.906.335.306.5021.9518.2725.6921.7216.9715.1214.9112.00
MajutsuCity8.007.838.608.007.007.507.407.8322.6220.9924.9919.9123.8920.5225.8422.70
Ours8.758.679.008.507.638.008.008.6727.8923.8327.6424.6529.4324.1129.5725.89

Table 2: Quantitative evaluation of building-level indoor–outdoor correspondence using GPT-5.5 and human evaluation. The best results are in bold. Dashes denote metrics unavailable for TRELLIS because its independently generated interior is not grounded in the exterior building footprint.

Table 2: Quantitative evaluation of building-level indoor-outdoor correspondence using GPT-5.5 and human evaluation. The best results are in bold. Dashes denote metrics unavailable for TRELLIS because its independently generated interior is not grounded in the exterior building footprint.
MethodAQSRDRShape IoU↑
Functional↑Visual↑Spatial↑Average↑Functional↑Visual↑
GPTHum.GPTHum.GPTHum.GPTHum.GPTHum.GPTHum.
TRELLIS6.236.305.884.8519.1720.031.8310.90
Ours7.477.458.217.608.778.658.157.9022.3520.3624.2923.510.997

Table 3: Effect of dynamic world-context localization on building-level indoor–outdoor correspondence under GPT-5.5 and human evaluation. The best results are in bold.

Table 3: Effect of dynamic world-context localization on building-level indoor-outdoor correspondence under GPT-5.5 and human evaluation. The best results are in bold.
ConfigurationAQSRDRShape IoU↑
Functional↑Visual↑Spatial↑Average↑Functional↑Visual↑Spatial↑
GPTHum.GPTHum.GPTHum.GPTHum.GPTHum.GPTHum.GPTHum.
Static World Context5.755.884.586.003.254.754.535.5416.3711.751.400.331.4011.090.670
Full Model7.677.888.178.257.758.507.868.2121.8218.8622.8518.8222.8517.500.994

Table 4: Effect of autoregressive neighborhood conditioning (ANC) on cross-block visual and spatial continuity under GPT-5.5 and human evaluation. The best results are in bold.

Table 4: Effect of autoregressive neighborhood conditioning (ANC) on cross-block visual and spatial continuity under GPT-5.5 and human evaluation. The best results are in bold.
ConfigurationContinuity AQS↑RDR↑
GPTHum.GPTHum.
Full Model8.257.7524.0421.97
w/o ANC7.255.3816.6619.07

Generated Worlds

Minecraft City

DemoComplete Minecraft City generated by HoloWorld
01 / 08FULL CITY OVERVIEW
Minecraft City local exterior rendered view 1
LOCAL VIEW · 01
Minecraft City local exterior rendered view 2
LOCAL VIEW · 02
Minecraft City interior dollhouse view
INTERIOR OVERVIEW
Minecraft City interior walkthrough view
INTERIOR VIEW

Cyberpunk City

DemoComplete Cyberpunk City generated by HoloWorld
02 / 08FULL CITY OVERVIEW
Cyberpunk City local exterior rendered view 1
LOCAL VIEW · 01
Cyberpunk City local exterior rendered view 2
LOCAL VIEW · 02
Cyberpunk City interior dollhouse view
INTERIOR OVERVIEW
Cyberpunk City interior walkthrough view
INTERIOR VIEW

European Old Town

DemoComplete European Old Town generated by HoloWorld
03 / 08FULL CITY OVERVIEW
European Old Town local exterior rendered view 1
LOCAL VIEW · 01
European Old Town local exterior rendered view 2
LOCAL VIEW · 02
European Old Town interior dollhouse view
INTERIOR OVERVIEW
European Old Town interior walkthrough view
INTERIOR VIEW

Ghibli City

DemoComplete Ghibli City generated by HoloWorld
04 / 08FULL CITY OVERVIEW
Ghibli City local exterior rendered view 1
LOCAL VIEW · 01
Ghibli City local exterior rendered view 2
LOCAL VIEW · 02
Ghibli City interior dollhouse view
INTERIOR OVERVIEW
Ghibli City interior walkthrough view
INTERIOR VIEW

High-Density Modern City

DemoComplete High-Density Modern City generated by HoloWorld
05 / 08FULL CITY OVERVIEW
High-Density Modern City local exterior rendered view 1
LOCAL VIEW · 01
High-Density Modern City local exterior rendered view 2
LOCAL VIEW · 02
High-Density Modern City interior dollhouse view
INTERIOR OVERVIEW
High-Density Modern City interior walkthrough view
INTERIOR VIEW

Candy-Pop Toy City

Complete Candy-Pop Toy City generated by HoloWorld
06 / 08FULL CITY OVERVIEW
Candy-Pop Toy City local exterior rendered view 1
LOCAL VIEW · 01
Candy-Pop Toy City local exterior rendered view 2
LOCAL VIEW · 02
Candy-Pop Toy City interior dollhouse view
INTERIOR OVERVIEW
Candy-Pop Toy City interior walkthrough view
INTERIOR VIEW

Modern City

Complete Modern City generated by HoloWorld
07 / 08FULL CITY OVERVIEW
Modern City local exterior rendered view 1
LOCAL VIEW · 01
Modern City local exterior rendered view 2
LOCAL VIEW · 02
Modern City interior dollhouse view
INTERIOR OVERVIEW
Modern City interior walkthrough view
INTERIOR VIEW

Oceanic Civilization

Complete Oceanic Civilization generated by HoloWorld
08 / 08FULL CITY OVERVIEW
Oceanic Civilization local exterior rendered view 1
LOCAL VIEW · 01
Oceanic Civilization local exterior rendered view 2
LOCAL VIEW · 02
Oceanic Civilization interior dollhouse view
INTERIOR OVERVIEW
Oceanic Civilization interior walkthrough view
INTERIOR VIEW

Click the “Demo” button to explore the interactive HoloWorld demo.

Citation

@article{huang2026holoworld,
  title={To See a World in a Living Context: Unified Indoor-Outdoor Urban World Generation},
  author={Huang, Xiaobin and Huang, Zilong and Luo, Yang and Fan, Hongchao and Chen, Yiping and Han, Ting},
  journal={arXiv preprint arXiv:2608.05879},
  year={2026}
}