Yo'City: Personalized and Boundless 3D Realistic City Scene Generation via Self-Critic Expansion

1. Introduction

• Prior city-generation approaches often rely on handcrafted rules, imagery, maps, or satellite data, and can struggle with flexible, user-friendly text inputs.SynCity’s flat tile-by-tile generation also lacks explicit urban hierarchy and can struggle with global consistency and realism at city scale.Source: §1 · p. 2
• Agentic 3D city generation remains largely unexplored, despite cities’ open, large-scale, highly structured spaces posing challenges for realistic and scalable synthesis.Manual construction is also challenging and labor-intensive, while existing generative approaches have limitations with flexible text inputs.Source: §1 · p. 2
• Yo’City uses a City–District–Grid planning strategy: a Global Planner allocates districts, then a Local Designer creates detailed grid-level descriptions.Planning proceeds from global structure to local architectural styles, building densities, landmarks, and surrounding context.Source: §1 · p. 2
• A produce–refine–evaluate loop creates isometric images for conversion into 3D assets, while scene-graph-guided expansion places new grids; the paper reports improvements across six evaluation dimensions.The benchmark dimensions include VQAScore, geometric fidelity, texture clarity, layout coherence, scene coverage, and overall realism.Source: §1 · p. 2

2.1. 3D City Generation

• Many 3D city methods build scenes from generated semantic maps or directly in 3D, but often require real-world data and offer limited user control for personalized, large-scale generation.Earlier procedural and image-based approaches can also be inefficient or inflexible.Source: §2.1 · p. 3

2.2. Agentic Systems

• LLM- and VLM-guided agents are used for indoor scene generation, but outdoor methods often rely on a single reference image and SynCity struggles with dense, realistic city-scale scenes.These approaches face challenges in generating vast urban environments with realism and fidelity.Source: §2.2 · p. 3

3.1. Problem Formulation

• Yo’City formulates city creation as planning, generation, and expansion, synthesizing all tiles in parallel from a global prompt rather than conditioning each tile on previously generated tiles.The parallel formulation avoids sequential dependency and potential error accumulation.Source: §3.1 · p. 3

3.2. Framework Overview

• The framework uses a Global Planner, Local Designer, 3D Generator, and Relationship-guided Expansion module to create and evolve a structured city scene.The expansion module selects a placement for a new tile, which is then synthesized and integrated.Source: §3.2 · p. 3

3.3. Global Planner

• The Global Planner estimates city size, identifies functional districts, and allocates them to a grid; retrieval-augmented information can ground plans for referenced real-world cities.District blueprints specify function and building types for subsequent local design.Source: §3.3 · pp. 3–4

3.4. Local Designer

• The Local Designer conditions on district blueprints and the global prompt to produce descriptions for all grids, jointly planning multi-grid districts for spatial and stylistic continuity.Its coarse-to-fine stages move from high-level organization to fine-grained grid detail.Source: §3.4 · p. 4

3.5. 3D Generator

• The 3D Generator produces isometric images on a shared ground platform, refines them, evaluates alignment, realism, and layout, then converts them to 3D models for grid-aligned assembly.The assembly adds roads and ground surfaces, and users can customize ground materials and attributes to match the city theme.Source: §3.5 · pp. 4–5

3.6. Relationship-guided Expansion

• For expansion, a VLM describes the requested grid and builds a scene graph of qualitative distance relations to existing districts; feasible locations are selected to satisfy those relations and preserve context.The stated spatial motivation includes coordinating nearby residential, school, and business areas while separating industrial districts.Source: §3.6 · p. 5
• Placement combines signed, distance-weighted spatial relations with Sentence-Bert similarity to neighboring grid descriptions, and minimizes the combined spatial and semantic objective.Positive relationship weights pull a candidate closer; negative weights push it away, while higher embedding similarity favors semantic compatibility.Source: §3.6 · p. 5
• After choosing the minimizing location, Yo’City generates the new grid’s 3D model and supports iterative, user-interactive city expansion.This mechanism is described as enabling open-world, boundless generation.Source: §3.6 · pp. 5–6

4.1. Settings

• The 100-prompt dataset combines 30% manually written and 70% GPT-4o-generated city descriptions; comparisons use Trellis, Hunyuan3D, CityCraft, and SynCity.Experiments use GPT-4o, GPT-Image-1, and Hunyuan3D (API) for language, image editing, and image-to-3D generation, respectively.Source: §4.1 · p. 6

4.2. Evaluation Metrics

• Visual quality is assessed through pairwise comparisons of geometric fidelity, texture clarity, layout coherence, scene coverage, and overall realism, alongside VQAScore for semantic consistency.GPT-5 and ten human judges each repeat comparisons twice.Source: §4.2 · p. 6

4.3. Main Results

• Yo’City achieves the highest VQAScore and win rates of at least 85.00% in Geometric Fidelity, 86.00% in Layout Coherence, 78.50% in Texture Clarity, and 84.50% in Overall Realism.These comparisons use the same input and pairwise visual-quality evaluations by GPT-5 and human judges.Source: §4.3 · p. 6
• Yo’City’s qualitative examples show coherent layouts, consistent scales, appropriate spacing, clear textures, and fine details; CityCraft struggles with non-modern prompts, while SynCity shows uneven density and coarse textures.The figure compares methods under the same city instructions and highlights SynCity regions with spatial inconsistency, lack of realism, and poor texture fidelity.Source: §4.3 · pp. 6–7, Figure 3 · p. 7
• At grid level, Yo’City exceeds SynCity by 0.0355 in Alignment Score and 0.57 in Aesthetic Score.Alignment Score uses VQAScore between a grid image and a prompt-based reasonableness query; Aesthetic Score is out of 10.Source: §4.3 · p. 7, Table 2 · p. 7

4.4. Ablation Studies

• With reasoning, Yo’City outperforms its no-reason variant in VQAScore, Layout Coherence, and Overall Realism, as evaluated by GPT-5.The comparison uses the same one-shot example; the authors attribute the improvement to the Global Planner and Local Designer capturing preferences and organizing layouts.Source: §4.4 · p. 8, Table 3 · p. 8
• Across successive expansions, VQAScores remain stable, with a coefficient of variation of 3.34%; a visualization shows five successive iterations and another example spans eight expansion steps.The expansion tasks evaluate VQAScore against each city’s global instruction.Source: §4.4 · p. 8, Figure 4 · p. 8

5. Conclusion

• Yo’City is a text-driven agentic framework for personalized, realistic 3D city generation without map data.Source: §5 · p. 8
• An LLM-based hierarchical planner designs city layouts, while a specialized generator creates isometric images and converts them into 3D models.Source: §5 · p. 8
• A relationship-guided mechanism supports iterative city growth through textual instructions.Source: §5 · p. 8
• Experiments show Yo’City outperforms existing methods across all dimensions, with potential applications in virtual reality and simulation games.Source: §5 · p. 8

A.1. Qualitative Comparison

• Across three city-instruction cases, Yo’City shows stronger fidelity and detail, with coordinated modern layouts and coherent Victorian facades and roofs.The results include four rendered viewpoints.Source: §A.1 · p. 12

A.2. Visualization of Expansion

• Across four expansion iterations, Yo’City preserves the original global instructions while adding coherent, visually consistent elements at locations selected using distance- and semantics-aware optimization.Source: §A.2 · p. 12

A.3. Expansion Mechanism Comparison

• Yo’City has the lowest VQAScore coefficient of variance among relationship-guided, random-placement, and semantic-only expansion strategies.Source: §A.3 · p. 12

B. Dataset Curation Details

• Human authors write 30% of the instruction dataset, and GPT-4o [1] generates the remaining 70%.Source: B. Dataset Curation Details · p. 12
• The dataset includes short sentences for concise requirements and long sentences for detailed scene descriptions and personalized intentions.Examples include city size or style references in short inputs and architectural or district requirements in long inputs.Source: B. Dataset Curation Details · pp. 12–13
• Sentence length is controlled to respect baseline input limits, although Yo’City can accommodate much longer inputs, including paragraph-level descriptions.Source: B. Dataset Curation Details · p. 13
• Keyword-style descriptions are included to accommodate varied user input habits.An example lists a modern city, entertainment hubs, high-rise buildings, and convenient life.Source: B. Dataset Curation Details · p. 13
• A prompt template is provided for generating diverse city instructions.Source: Figure 6 · p. 12, B. Dataset Curation Details · p. 13

C. Implementation Details

• Experiments run on Ubuntu 22.04 with an Intel Xeon E5-2699 v4 CPU and NVIDIA A6000 GPUs with 48GB memory.Source: C. Implementation Details · pp. 13–14
• GPT-4o and GPT-Image-1 are accessed through official APIs; image refinement runs up to three iterations and accepts evaluator scores of 6 or higher.Below 6, the VLM rewrites the instruction using negative feedback; the loop stops when an acceptable score is reached.Source: C. Implementation Details · p. 14
• Refinement removes background content and tile structures to prevent visible grid seams; image evaluation indirectly promotes cross-grid consistency by checking descriptions.Source: C. Implementation Details · p. 14
• Post-processing scales grid models and integrates ground and roads; expansion regularization uses Sentence-Bert cosine similarity and relationship weights from 1 for near to -1 for far.The regularization parameter λ is set to 1.Source: C. Implementation Details · p. 14

• VQAScore measures semantic consistency by aligning each generated city image with its input text.The alignment uses the instruction plus a prompt requesting balanced proportions and realistic, non-exaggerated forms; images use identical viewpoints.Source: D. Evaluation Details · p. 14
• Visual quality is assessed across geometric fidelity, texture clarity, layout coherence, scene coverage, and overall realism.GPT-5 and ten human judges make pairwise comparisons for each aspect, repeating each comparison twice.Source: Figure 9 · p. 15, D. Evaluation Details · p. 15
• Grid-level alignment uses VQAScore to test whether each grid reasonably reflects the city instruction.A separate grid-level aesthetic score uses a SigLIP predictor, with scores of 5.5+ regarded as great.Source: D. Evaluation Details · pp. 15–16
• Table 5 covers runtime, GPU memory, and VQAScore.Source: Table 5 · p. 16
• Figure 10 compares generation time across city sizes and reports that Yo’City is consistently more efficient than SynCity.Source: Figure 10 · p. 16

E.1. Computational Efficiency

• The comparison measures generation time and GPU memory against CityCraft and SynCity, with API substitutions for a fair memory comparison.All three methods access GPT-4o via API; Yo’City’s image and 3D generation APIs are replaced with FLUX.2-Klein-9B and Trellis for the memory comparison.Source: §E.1 · p. 16
• Yo’City’s non-autoregressive generation allows different city areas to be generated in parallel, unlike SynCity’s sequential process, and avoids complex blending.Source: §E.1 · p. 16
• 43.40 min for a typical 3×3 city without parallelization is only 69.47% of SynCity’s runtime.Figure 10 compares time consumption across city sizes, with Yo’City consistently requiring less time than SynCity.Source: §E.1 · p. 16

Limitations and Future Work

• Using pretrained models for inference reduces data dependence and allows more flexible inputs, but Yo’City’s performance may be limited by their capabilities.Source: §E.2 · p. 16
• The authors propose continually integrating cutting-edge methods to enhance Yo’City’s competence.Source: §E.2 · p. 16
• The current model focuses on city structure and infrastructure, omitting natural surroundings such as mountains and seas; future work could incorporate these features.Source: §E.2 · p. 16

Expansion prompt

• The prompt asks the designer to analyze the city image, existing zones, and user request before describing a new block that harmonizes with the layout.It considers architectural logic, density, and functional organization, as well as the requested block and grid position.Source: §E.2 · p. 22

Architectural constraints

• Block descriptions focus on buildings, massing, facades, courtyards, and layout, emphasizing proportional balance, material consistency, and spatial continuity.The prompt excludes people, vehicles, atmosphere, and excessive greenery.Source: §E.2 · p. 22

Spatial relations

• Spatial relations use conceptual proximity terms and include only meaningful, logically consistent adjacencies; no edge is generated when a clear relationship is absent.“Far” is reserved for explicit spatial separation rather than applied broadly.Source: §E.2 · p. 22