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 framework for generating personalized, infinitely expandable 3D city scenes from free-form text prompts. It uses a hierarchy of LLM/VLM agents for top-down planning, parallel grid generation, and scene-graph-guided spatial expansion—without requiring map or satellite training data.
Problem¶
Existing 3D city generation methods either rely on GAN/diffusion models trained on geographic data (limiting text-based personalization and scalability) or use flat tile-by-tile auto-regressive pipelines (e.g., SynCity) that lack hierarchical spatial reasoning, producing cartoonish geometry, blurry textures, and poor global consistency at city scale.
Method¶
Yo'City decomposes generation into four agent modules:
- Global Planner: An LLM estimates city size (HĂ—W grid), identifies N functional districts, and allocates them spatially. A RAG module retrieves Wikipedia data about real-world reference cities (distilled via GPT-4o-mini) to ground structural priors.
- Local Designer: Per-district LLM refines Global Planner blueprints into detailed per-grid textual descriptions, reasoning over all grids within a district jointly for stylistic coherence.
- 3D Generator: All grids are generated in parallel (breaking SynCity's causal tile dependency) via a produce–refine–evaluate loop: (i) generate isometric image on a ground-plane anchor, (ii) edit to remove artifacts and enhance diversity, (iii) VLM evaluator checks text-image alignment, realism, and layout rationality; iterate until criteria are met. Approved isometric images are lifted to 3D via a pretrained image-to-3D model, then assembled using the predefined layout.
- Expansion Module: Given a user expansion request, a VLM generates a description for the new grid (
d_new) and constructs a scene graph encoding qualitative distance relationships (near/far) to existing districts. An optimization objective minimizes a weighted Euclidean distance term (with signed weights per relationship type) plus a Sentence-BERT semantic regularization term, selecting the placement that is spatially and contextually coherent.
Key Contributions¶
- Multi-agent "City–District–Grid" hierarchical framework enabling text-driven, boundless 3D city generation without task-specific training data.
- RAG-augmented Global Planner that grounds layout decisions in real-world city structure from Wikipedia.
- Parallel grid generation pipeline that eliminates sequential error accumulation from tile-by-tile approaches.
- Scene-graph-based expansion with distance- and semantics-aware placement optimization (combined
L_dist+L_semobjective). - Multi-dimensional evaluation benchmark covering VQAScore, geometric fidelity, layout coherence, texture clarity, scene coverage, and overall realism.
Results¶
- Yo'City outperforms SynCity and other state-of-the-art methods across all six evaluation dimensions: VQAScore (semantic alignment), geometric fidelity, layout coherence, texture clarity, scene coverage, and overall realism.
- Specific numerical scores are not provided in the available text excerpt.
Limitations¶
- Relies on an off-the-shelf pretrained image-to-3D model as a black box; quality ceiling is bounded by that model's fidelity.
- RAG corpus is limited to Wikipedia, potentially insufficient for highly niche or fictional city references.
- Expansion module uses breadth-first search over candidate locations, which may not scale efficiently to very large existing cities.
- No stated upper bound on city size or compute cost for large HĂ—W grids.
- Evaluation benchmark is self-constructed; no comparison to externally validated 3D scene quality metrics.
Relevance to Agentic AI / LLM Agents¶
Yo'City exemplifies compositional multi-agent system design: distinct LLM/VLM agents handle planning, refinement, evaluation, and expansion—each with a scoped role and tool use (RAG, image generation, image editing, scene graph construction). The produce–refine–evaluate loop is a concrete instance of agentic self-critique and iterative improvement applied to a non-text output domain. The scene-graph-based expansion module shows how agents can maintain persistent world state and reason spatially about it, a capability directly relevant to embodied agents and long-horizon planning. This work extends the LLM-agent-for-3D-scene-generation line (previously focused on indoor scenes) to open, large-scale outdoor environments.