A new AI model designs brick structures that are actually possible to build, not just to render.
Researchers introduced BrickAnything, an AI system that takes a 3D shape - fed in as a point cloud - and outputs a sequence of bricks that reconstruct it while obeying real assembly constraints. Instead of placing each brick independently, the system uses what its authors call structure-aware tree tokenization, which encodes how each brick attaches to the ones around it. That keeps the generation process closer to how someone would actually stack bricks, and cuts down on invalid in-progress states. The team also added preference-based training, decoding rules that block invalid moves, and an adaptive rollback mechanism so the model can undo a bad placement instead of shipping a structure that would collapse.
Most existing brick-generation tools lean on heuristic optimization, which breaks down when a shape simply cannot be built under the rules, or generate brick sequences without modeling the underlying geometry and assembly relations at all. BrickAnything's bet is that baking structural relationships directly into generation, rather than bolting on stability checks afterward, produces results that are both geometrically faithful and physically realizable. That distinction matters for anyone trying to push 3D generative AI past pretty renders and into things people can actually assemble.
It's a research paper, not a product, and arXiv preprints have a habit of staying preprints - but buildable-by-construction generation is a more useful trick than one more photorealistic 3D model nobody can print, pour, or snap together.