Overview
- Field: Computer Vision
- Authors: Peter Kulits, Cordelia Schmid
- Published: 2026-11-12
- arXiv: 2611.17321
- Task: assemble a set of LEGO parts to match a single reference photo — a discrete, combinatorial alternative to continuous 3D reconstruction.
- Method: a set-conditioned SE(3) flow model pretrained on procedurally generated assemblies.
- Data-efficient finetuning: image conditioning is added with only 1% of the data.
- Novelty: analytic connectivity-field guidance at test time ensures predictions form connected assemblies.
- Findings: model scale, pretraining, and connectivity guidance jointly improve construction quality.
Abstract
Given a set of LEGO parts and an image of an assembled object, the task is to infer the pose of each part to assemble the object. In contrast with traditional 3D reconstruction, doing so requires satisfying discrete physical constraints, which provide structure but render the problem combinatorial and highly sensitive to small pose errors.
To learn a prior over how parts fit together, the authors train a set-conditioned SE(3) flow model on procedurally generated assemblies. They then finetune this model to inject image conditioning, using 1% as much data.
While the learned flow captures global structure, it does not enforce connectivity. To address this, the paper introduces an analytic connectivity-field guidance that pulls predictions toward connected assemblies at test time.
Key points
Resources
Model and data are openly available at: https://kulits.github.io/BrickFlow/
--- *Auto-collected on 2026-11-14.*