BrickAnything: Geometry-Conditioned Buildable Brick Generation with Structure-Aware Tokenization

BrickAnything is a new AI framework that generates physically buildable brick structures from 3D shapes by combining geometric reconstruction with structural constraints. The method uses structure-aware tokenization to model how bricks attach to each other, improving the feasibility and stability of generated designs compared to existing heuristic approaches.

BrickAnything addresses a computational challenge at the intersection of 3D geometry and physical constraints: automatically generating brick structures that are both geometrically accurate and physically constructible. Traditional approaches rely on heuristic optimization, which frequently fails when target shapes lack feasible brick-based solutions under predefined constraints. This new framework reframes the problem as a sequence generation task, treating brick assembly as an autoregressive process that mirrors how physical construction actually occurs.

The innovation centers on structure-aware tree tokenization, which represents bricks through their local attachment relations rather than arbitrary ordering. This approach fundamentally aligns the AI model's decision-making process with physical reality, reducing invalid intermediate states that would require backtracking. By using point clouds as a unified geometric interface, the system achieves flexibility across diverse 3D input formats while maintaining explicit modeling of assembly dependencies.

The broader significance extends beyond brick generation to procedural content creation, architectural design automation, and robotics applications. This represents a maturation in how AI systems handle constrained generation problems where outputs must satisfy discrete rules and physical laws simultaneously. The incorporation of preference-based alignment post-training and validity-constrained decoding demonstrates emerging best practices for aligning learned models with real-world requirements.

For AI researchers and developers, this work establishes patterns for handling multi-objective generation tasks where geometric fidelity and structural feasibility matter equally. Future applications could include automated toy design, architectural planning tools, and robotic assembly systems. The methodology's success in reducing rollback and regeneration suggests potential efficiency gains in other constrained sequence generation domains.

• →BrickAnything generates physically buildable brick structures using autoregressive AI guided by structure-aware tokenization of assembly relations.
• →The framework addresses limitations of heuristic optimization by explicitly modeling geometric constraints and physical construction sequences.
• →Structure-aware tokenization reduces invalid intermediate states by aligning AI decisions with actual brick attachment mechanics.
• →Post-training techniques including preference-based alignment and validity-constrained decoding improve both stability and geometric fidelity.
• →The approach demonstrates how discrete constraint satisfaction can be integrated into modern generative AI architectures.