When we imagine a building or a geological formation, we don't think in terms of a grid of colored pixels or a cloud of points. We think in terms of boundaries, paths, and structural relationships. However, for years, generative models for spatial environments relied heavily on rasterized representations—images or occupancy grids—which often struggle to maintain basic topological rules like perfectly straight walls or continuous riverbeds.
A growing body of research suggests that the key to generating professional-grade layouts lies in treating geometry not as a picture, but as a sequence of vectors. By representing environments as collections of line segments and space vectors, models can preserve the inherent structural logic of a scene while leveraging the same autoregressive power that drives modern language models.
The Vector Advantage over Raster Grids
The primary challenge in spatial generation is "structural integrity." In a pixel-based model, a wall is just a line of dark pixels. If the model predicts a few pixels incorrectly, the wall might appear "fuzzy" or, worse, have a gap that breaks the room's topology.
Jiachen Liu and colleagues in End-to-End Graph-Constrained Vectorized Floorplan Generation with Panoptic Refinement argue that vectorized representations are fundamentally superior for architectural tasks. By generating floorplans as vectors directly from room connectivity graphs, their framework ensures that walls remain straight and room volumes stay connected. This approach avoids the "phantom obstacles" and topological errors common in image-based methods, producing layouts that are immediately editable in professional CAD software.
This principle extends beyond human architecture into natural structures. In geological modeling, Chao Luo et al. demonstrate in Modeling of dense well block point bar architecture based on geological vector information that using space vectors to define the curvature of ancient river channels allows for the automated generation of internal reservoir architectures. Because the model follows a vector path rather than a grid, the internal layers of the deposit naturally follow the primary depositional trend, maintaining geological consistency that a voxel-based approach would likely break.
Geometry as a Sequence Task
If we accept that vectors are the right primitive, the next question is how to generate them. Recent breakthroughs have reframed geometric generation as a sequence modeling problem, similar to how Large Language Models (LLMs) predict the next word in a sentence.
In Beyond the Frontier: Predicting Unseen Walls From Occupancy Grids by Learning From Floor Plans, Ludvig Ericson introduces Floorist, a transformer-based model that treats floor plan prediction as a sequence of vector line segments. Instead of predicting a whole image at once, the model generates "vector tokens"—coordinates that define the start and end of wall segments.
This autoregressive approach allows the model to reason about architectural dependencies. For example, if the model "sees" a doorway, it can logically predict the walls that should follow it. By treating geometry as a language of segments, the system can "hallucinate" unobserved parts of a building with far greater reliability than traditional regression models, which often produce noisy or physically impossible predictions in unknown spaces.
Why It Works: Inductive Biases
The success of these methods stems from a better alignment between the model's "inductive bias" and the reality of the physical world:
- Sparsity and Efficiency: A complex floorplan can be described by a few dozen vectors, whereas a high-resolution image requires millions of pixels. This makes the learning task more computationally efficient.
- Topological Constraints: Vectors inherently encode connectivity. It is easier for a model to learn that two points are connected than to learn that a string of pixels should appear continuous.
- Contextual Reasoning: By using transformers to process these vectors, models can capture long-range dependencies—ensuring that a wall on one side of a building aligns perfectly with a wall on the other.
By moving from pixels to vectors, researchers are enabling machines to understand and generate the world not just as a collection of shapes, but as a coherent system of structural rules.