TL;DR
The paper introduces Complexity-Diversity Building Metrics (CDBM), a set of 12 interpretable descriptors that measure both the intensity of geometric detail and its distribution across building massings. By analyzing over 482 million buildings, the authors demonstrate that these metrics are superior for distinguishing real from AI-generated architecture and align more closely with human aesthetic preferences than existing "black-box" or purely volumetric methods.
What Problem Does This Paper Address?
In architecture and urban planning, "massing"—the basic 3D shape of a building—is the foundation of design. With the rise of generative AI, we can now create millions of building massings automatically. However, we lack a good way to evaluate them.
Existing metrics are often too simplistic (focusing only on footprint area or height) or too opaque (using deep learning "latent spaces" that provide a score but no architectural explanation). There is a specific gap in measuring geometric detailing: how do we quantify the difference between a simple box and a complex, heterogeneous structure with varied setbacks and irregular angles? The authors aim to provide a "cost-efficient, interpretable descriptor" that can diagnose why a generated building looks "fake" or unappealing.
Methods
The core of the paper is the Complexity-Diversity framework. It breaks down building geometry along two axes:
- Complexity vs. Diversity:
- Complexity measures the fineness of detail (e.g., how many short edges or sharp, non-orthogonal angles a building has).
- Diversity measures heterogeneity (e.g., does the building mix long and short walls, or are they all the same?).
- Linear vs. Planar:
- Linear properties look at the 1D "skeleton" (edge lengths, vertex angles, and extrusion heights).
- Planar properties look at the 2D footprint shape (circularity, convexity, and elongation).
Crossing these axes results in 12 distinct metrics (e.g., "Side Length Complexity," "Extrusion Height Diversity"). These are calculated in "closed form," meaning they are mathematically precise and computationally cheap enough to run on hundreds of millions of buildings.
Datasets
The authors didn't just propose a theory; they tested it at a massive scale:
- OSM-derived Dataset: They processed 482,241,306 buildings from OpenStreetMap into a structured "Level of Detail (LoD) 1.3" format (stacked polygons with specific heights).
- Site-Massing Dataset: 15.7 million sites paired with their building massings, used for training and validation.
- Generative Benchmark: They created a new benchmark of 12,000 generation tasks to see if AI models (like Qwen and TRELLIS) could hit specific geometric targets defined by the CDBM metrics.
Key Insights and Ideas
- AI Models are "Geometric Minimalists": One of the most striking findings is that current generative models struggle to produce high-complexity and high-diversity forms. They are quite good at making "average" buildings but fail to generate the "strongly expressed and heterogeneous forms" found in real-world architecture.
- Interpretability Matters for Design: Unlike deep learning latents (like LION), CDBM can tell a designer exactly what is wrong. For example, a model might be failing because it produces uniform heights (low Extrusion Height Diversity) compared to its neighbors.
- Human Preference Alignment: The researchers found that a classifier trained on CDBM metrics was a better proxy for human visual preference than traditional 3D metrics. This suggests that the "complexity" and "diversity" we see in buildings are key components of what we perceive as "realistic" or "good" design.
Key Results
- Real vs. Generated Discrimination: Using the CatBoost classifier, CDBM achieved an ROC-AUC of 0.972 in separating real buildings from AI-generated ones, significantly outperforming the 3DBM baseline (0.931) and the LION latent representation (0.837).
- Human Preference: CDBM showed the strongest positive correlation with human rankings of building "realism" across four different AI generators.
- Contextual Relevance: While building orientation remains the strongest predictor of whether a building "fits" its neighborhood, CDBM metrics provided complementary information that improved the accuracy of contextual fit models.
Limitations and Future Work
The authors acknowledge several constraints:
- LoD 1.3 Limitation: The metrics currently only work on "block models" (stacked extrusions). They don't yet account for pitched roofs, facade textures, or finer architectural details.
- Single-Building Focus: Most of the experiments focused on single-building sites. Real urban environments are multi-building complexes, which the authors plan to address in future work.
- Classifier Dependency: The performance of these metrics depends heavily on the type of classifier used (non-linear models like CatBoost performed much better than simple logistic regression).
In conclusion, CDBM offers a powerful, interpretable "ruler" for the next generation of AI-driven urban design, helping us move past simple boxes toward more diverse and complex architectural forms.
Summary updated