Most AI room-layout systems fall apart the moment walls stop meeting at 90 degrees.
Researchers have introduced SPG-Layout, a text-driven framework for generating 3D indoor scenes in what the field calls non-Manhattan environments — rooms with angled walls, irregular footprints, or any geometry that defies the tidy grid most layout models assume. The system works in two main ways: it draws on statistical priors of how objects are typically distributed in a space to sharpen its sense of physical plausibility, and it borrows from human interior-design practice by placing large objects first, then arranging smaller ones around them. The team also built a benchmark of 500 diverse non-rectangular environments to test against — a data set that didn't exist before this work. In experiments, SPG-Layout outperformed existing methods on both standard rectangular and irregular layouts, and the code is slated for public release.
The gap this addresses is more practical than it might sound. Real floor plans — renovated apartments, converted warehouses, older buildings in dense cities — routinely ignore the right-angle assumption baked into most generative tools. A system that breaks on anything but a grid is a system that breaks constantly in the real world. What's notable here is the hierarchical strategy: prioritizing big furniture first is how human designers actually work, and encoding that heuristic into the model is a meaningful departure from purely geometric optimization.
LLM-driven scene synthesis has moved fast, but most published benchmarks quietly test only box-shaped rooms — a convenient shortcut that overstates how well these tools would perform on actual architecture.