Research map / 3D vision / Face and shape

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

NeurIPS 2026 · 3D vision › Face and shape · approach: 3D shape generative modeling · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes BrickAnything, a point-cloud-conditioned autoregressive model that generates buildable LEGO-style brick structures using a tree-based attachment tokenization, DPO post-training, validity-constrained decoding, and stability-guided rollback.

Problem

Generating physically buildable brick structures that faithfully reconstruct a 3D shape while satisfying discrete part constraints and structural stability; heuristic methods fail when no feasible assembly exists and existing learned generators are conditioned on text or images rather than explicit

Approach

Replace coordinate-ordered brick sequences with a BFS spanning-tree tokenization that encodes each child brick relative to its parent's stud (parent-side connector, child size, child-side anchor), condition on point clouds via Michelangelo, and align with a buildability reward (voxel IoU + Chamfer + min per-brick stability) using reward-weighted DPO, plus parent-aware rollback at inference.

Key results

• On the challenging subset, BrickAnything reaches 83.4% stable rate, 100% valid rate, 0.586 IoU and 0.422 average rollbacks, versus 76.0% / 0.553 / 6.750 for the BrickGPT-style baseline with DPO.
• On the stable subset, IoU improves from 0.742 (BrickGPT-style) to 0.788 and average rollbacks drop from 3.620 to 0.184, with 100% stable and valid rates.
• Legolization obtains 0.0% stable rate on the challenging subset where it fails, while the proposed method is learned and does not rely on heuristic search.

Surprising findings

• Removing validity-constrained decoding, rollback and DPO gives the highest IoU on the challenging subset (0.602) yet only 60.6% stable rate, showing geometric overlap does not imply buildability.

Limitations

• Restricted 20x20x20 workspace with only eight standard brick types limits resolution and geometric complexity.
• Failures on slender components, smooth curved surfaces, and fine local details.
• Buildability reward is hand-designed; training data is generated by Legolization rather than human-designed brick models.

Experimental setup

Datasets: ShapeNet Objaverse Objaverse-XL

Benchmarks: stable subset (500 samples) challenging subset (500 samples)

Metrics: Chamfer distance voxel IoU % stable % valid average rollback count

Baselines: Legolization BrickGPT-style tokenization BrickGPT-style tokenization w/o DPO

Builds on: BrickGPT StableLego Legolization Michelangelo OPT DPO

Closest papers by meaning

• SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation (NeurIPS 2026 · similarity 0.86)
• StepCAD: Mesh-to-CAD Code Generation via LLM Policy and Geometry-Guided Search (NeurIPS 2026 · similarity 0.85)
• Enhancing Implicit Shape Generators Using Topological Regularizations (ICML 2024 · similarity 0.84)
• DecompDreamer: A Composition-Aware Curriculum for Structured 3D Asset Generation (NeurIPS 2026 · similarity 0.84)
• CADMorph: Geometry‑Driven Parametric CAD Editing via a Plan–Generate–Verify Loop (NeurIPS 2025 · similarity 0.83)
• ArcPro: Architectural Programs for Structured 3D Abstraction of Sparse Points (CVPR 2025 · similarity 0.83)