Reproduce “Constellation: Learning relational abstractions over objects for compositional imagination” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
James C.R. Whittington, Rishabh Kabra, Loic Matthey, Christopher P. Burgess, Alexander Lerchner
Constellation learns scene arrangements separately from object appearance, allowing layouts to change while retaining the objects. It uses pre-trained MONet object slots and a relational encoder to represent group configuration, then applies symbolic labels such as “circle” to alter relational latents. The authors report five separate relational factors and demonstrate label-conditioned rearrangements on synthetic scenes, not complex real-world scenes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.