Reproduce “Long-LRM: Long-sequence Large Reconstruction Model for Wide-coverage Gaussian Splats” 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

Chen Ziwen, Hao Tan, Kai Zhang, Sai Bi, Fujun Luan, Yicong Hong, Li Fuxin, Zexiang Xu

Long-LRM introduces a feed-forward model for instant, wide-coverage 3D Gaussian Splatting reconstruction from numerous high-resolution images. The system achieves an 800x speedup over optimization-based methods by reconstructing a scene in approximately 1 second and delivers rendering quality comparable to or surpassing these slower approaches, demonstrating robust generalizability.

Key insights

  • A hybrid Mamba2-Transformer architecture effectively balances efficient long-context reasoning with global context preservation, making it suitable for handling extremely long sequences in complex visual tasks.
  • Token merging is a critical component for scaling generalizable 3D reconstruction models to ultra-high resolution and numerous input images, enabling practical wide-coverage scene synthesis without prohibitive memory or computational costs.
  • The combination of scale-invariant depth regularization and L1 opacity regularization significantly improves reconstruction stability, reduces artifacts like 'floaters,' and enables efficient pruning of redundant Gaussians during inference.

Summary by alphaXiv.

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