TriSplat: Simulation-Ready Feed-Forward 3D Scene Reconstruction

TriSplat is a feed-forward 3D reconstruction method that turns images into simulation-ready triangle meshes in one pass.

The network predicts oriented triangle primitives, camera poses, and optional intrinsics, using point maps and normals to stabilize geometry.

Experiments on RealEstate10K and DL3DV show better geometric fidelity than Gaussian-based baselines, with competitive rendering quality.

The result narrows the gap between image-based reconstruction and downstream use in physics engines, collision detection, and embodied systems.