BrickAnything: AI Generates Buildable Brick Structures from 3D Shapes
A new framework called BrickAnything has been unveiled by researchers, designed to generate brick structures that can be physically constructed from various 3D representations. This innovative method employs point clouds as a comprehensive geometric interface, predicting sequences of bricks that reconstruct the desired shape while adhering to assembly constraints. To capture the structural relationships between bricks, the team created a structure-aware tree tokenization that illustrates brick configurations through local attachments. This technique overcomes the shortcomings of current methods reliant on heuristic optimization, which may struggle when a target 3D shape cannot be feasibly structured under set constraints or fails to accurately model the essential 3D geometry and assembly relationships. The research is documented in a paper on arXiv with ID 2605.26182.
Key facts
- BrickAnything is a geometry-conditioned autoregressive framework for generating buildable brick structures.
- It uses point clouds as a unified geometric interface.
- The framework predicts brick sequences that reconstruct target shapes under assembly constraints.
- A structure-aware tree tokenization represents brick structures through local attachments.
- Existing methods rely on heuristic optimization that can break down for infeasible shapes.
- Other methods generate brick sequences without modeling 3D geometry or assembly relations.
- The paper is available on arXiv with ID 2605.26182.
- The approach addresses both discrete part constraints and structural stability.
Entities
Institutions
- arXiv