Reproduce “LeafFit: Plant Assets Creation from 3D Gaussian Splatting” 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

Chang Luo, Nobuyuki Umetani

LeafFit converts high-fidelity 3D Gaussian Splatting (3DGS) captures of individual plants into lightweight, editable, and instanced mesh assets suitable for real-time applications. The pipeline reduces a typical plant's mesh size to 1.13 MB with 2432 vertices, achieving 11,980.11 frames per second, while accurately preserving thin leaf details and enabling efficient artistic editing.

Key insights

  • Plant-specific leaf segmentation can be achieved effectively by computing geodesic distances from a root Gaussian, allowing for robust isolation of individual leaves and petioles without prior training.
  • Utilizing differentiable MLS for non-rigid registration enables accurate deformation of a single template to diverse leaf shapes, facilitating efficient GPU instancing and memory reduction for complex foliage.
  • Reconstructing thin meshes with the Ball Pivoting Algorithm directly from downsampled Gaussians, rather than implicit surfaces, preserves fine details and thinness, leading to high-quality, game-ready assets compatible with standard rendering pipelines.

Summary by alphaXiv.

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