Reproduce “LIM: Large Interpolator Model for Dynamic Reconstruction” 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

Remy Sabathier, Niloy J. Mitra, David Novotny

The Large Interpolation Model (LIM) is a transformer-based feed-forward framework that interpolates 3D implicit representations between video keyframes to reconstruct dynamic 4D assets. This approach generates traceable, consistently UV-textured meshes at high speeds, demonstrating superior temporal consistency and significantly faster processing for dynamic content from monocular video compared to prior optimization-based methods.

Key insights

  • Extending a static large reconstruction model (LRM) to dynamic 4D interpolation through a feed-forward, transformer-based design drastically improves reconstruction speed.
  • A self-supervised causal consistency loss is critical for ensuring smooth, continuous-time interpolation and strong temporal coherence in dynamic 4D asset generation.
  • Implicitly encoding canonical surface coordinates within the 3D representation provides an effective mechanism for accurate, topology-preserving mesh tracking, yielding production-ready deformable assets.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.