Skip to content

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

🕒 Published (v1): 2025-11-24 04:02 UTC · Source: Arxiv · Venue: CVPR 2026 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Yo'City is a multi-agent agentic framework for boundless, text-driven 3D city generation that replaces single-model diffusion pipelines with a hierarchical "City–District–Grid" planning strategy, a produce–refine–evaluate isometric image loop, and a scene-graph-guided self-critic expansion module. All grid tiles are generated in parallel rather than autoregressively, eliminating error accumulation. The system outperforms prior methods across six evaluation dimensions on a newly constructed benchmark.

Problem

Existing 3D city generation methods train a single diffusion model conditioned on maps or satellite data, which limits personalization, scalability, and user control. The closest training-free alternative (SynCity) uses a sequential tile-by-tile pipeline without hierarchical planning, causing error accumulation and poor global consistency at city scale. Neither class of methods supports interactive, unbounded city expansion guided by spatial logic.

Method

Yo'City decomposes generation into four coordinated modules:

  1. Global Planner (LLM): Takes a free-text user prompt \(p_0\) and outputs a city layout of \(H \times W\) grid cells, estimating city size, defining \(N\) functional districts with blueprints \(\{B_i\}\), and allocating them spatially. A RAG module over a Wikipedia corpus grounds references to real cities (e.g., "New York-like").

  2. Local Designer (LLM): For each district blueprint \(B_i\), generates fine-grained per-grid descriptions \(\{d_i\}_{i=1}^{H \times W}\) covering architectural style, density, and landmarks. All grids in a district are planned jointly to enforce intra-district coherence.

  3. 3D Generator: Each \(d_i\) drives an iterative produce–refine–evaluate loop: (i) generate an isometric image on a fixed ground platform anchor; (ii) an image-editing model removes the platform and corrects artifacts; (iii) a VLM evaluator checks text-image alignment, realism, and layout rationality—looping until criteria pass. Approved images are lifted to 3D assets via a pretrained image-to-3D model. All grids are generated in parallel, then assembled without complex 3D blending.

  4. Relationship-guided Expansion Module: On user expansion demand, a VLM generates a description \(d_\text{new}\) for the new grid and constructs a scene graph encoding qualitative distance relations (near/far) to existing districts. The optimal placement \(x^*\) is found by minimizing: $\(L(x) = \sum_{g \in \mathcal{G}} \gamma_{r(g)} \|x - g\|_2 - \lambda \sum_{y \in \mathcal{N}(x)} \text{Embedding\_Sim}(d_\text{new}, d_y)\)$ where \(\gamma_{r(g)}\) is signed (positive = pull closer, negative = push away) and the second term is a Sentence-BERT semantic regularizer encouraging neighborhood compatibility.

Key Contributions

  • Multi-agent agentic framework for personalized, boundless 3D city generation from free-text prompts, eliminating training on map/satellite data.
  • Hierarchical coarse-to-fine "City–District–Grid" planning with a RAG module for real-world city grounding; parallel (not sequential) grid generation avoids error accumulation.
  • Produce–refine–evaluate isometric image synthesis loop with a VLM-based quality evaluator as an inline critic.
  • Scene-graph-based, distance- and semantics-aware expansion objective for spatially coherent city growth.
  • Multi-dimensional evaluation benchmark covering VQAScore, geometric fidelity, texture clarity, layout coherence, scene coverage, and overall realism.

Results

  • Yo'City outperforms prior state-of-the-art methods across all six evaluation dimensions on the authors' constructed benchmark.
  • Baselines include SynCity (tile-by-tile pipeline) and other prior 3D city generation methods.
  • Specific numerical scores are not reported in the provided text excerpt; the paper states consistent superiority on VQAScore [31], geometric fidelity, layout coherence, texture clarity, scene coverage, and overall realism.

Limitations

  • Benchmark is constructed by the authors; no evaluation on an independently established external benchmark.
  • Specific numerical results not included in the provided text, making direct quantitative comparison difficult.
  • The expansion objective (Eq. 3) requires tuning of \(\lambda\) and qualitative relationship weights \(\gamma_{r(g)}\); sensitivity analysis is not discussed in the excerpt.
  • System depends on multiple external pretrained models (GPT-4o-mini, image-to-3D, Sentence-BERT, VLM evaluator); quality is contingent on their capabilities.
  • Computational cost of the per-grid produce–refine–evaluate loop across large \(H \times W\) grids is not analyzed.

Relevance to Harnesses / Meta-Harnesses

Yo'City is a concrete instantiation of a multi-agent meta-harness: it orchestrates a heterogeneous set of specialized models (LLM planner, LLM designer, image generator, image-editing model, VLM evaluator, image-to-3D model, Sentence-BERT) through a structured pipeline with feedback. The produce–refine–evaluate loop is a self-correcting inner harness where a critic model gates forward progress—a pattern directly analogous to harness designs that wrap generation with automated quality filters before committing output. The scene-graph-based expansion module demonstrates how a meta-harness can maintain persistent world state (the city graph) across iterative agent invocations, enabling unbounded, stateful generation rather than one-shot output. For researchers building harnesses, this paper offers a working example of hierarchical decomposition (global → local → grid), parallel agent dispatch with a shared layout contract, and VLM-as-judge integration within an automated loop.