BrickFlow: Connectivity-Guided Brick Reconstruction
BrickFlow adds test-time guidance to steer predicted LEGO part placements toward connected assemblies.
Given a set of parts and an image of the finished object, the model predicts each part’s position and orientation. It learns how parts fit together from procedurally generated assemblies, then gains image conditioning through finetuning with 1% as much data. Because the learned model captures overall structure without enforcing connectivity, an analytic connectivity field guides its predictions at test time. The authors report improved build quality with larger models, pretraining and guidance, and make models and data available for research.
HF Daily Papers' note
Given a set of parts and an image of the finished object, the model predicts each part’s position and orientation. It learns how parts fit together from procedurally generated assemblies, then gains image conditioning through finetuning with 1% as much data. Because the learned model captures overall structure without enforcing connectivity, an analytic connectivity field guides its predictions at test time. The authors report improved build quality with larger models, pretraining and guidance, and make models and data available for research.
HF Daily Papers' note
score 4