TriSplat is a feed-forward network that reconstructs a 3D scene from sparse, unposed images and exports it as a simulation-ready triangle mesh — in a single forward pass. No camera calibration, no post-hoc mesh extraction step.
## Triangles instead of Gaussians
Most recent reconstruction pipelines use Gaussian splatting, which renders beautifully but needs a slow, separate stage to convert into the mesh a physics engine actually wants. TriSplat skips that by predicting oriented triangle primitives directly, alongside camera poses, point maps, and appearance attributes. Because the output is already triangles, it loads straight into physics engines, collision detectors, and standard renderers.
## Why it matters
The speed gap is the story: TriSplat exports a simulation-ready mesh in under 1.3 seconds, where Gaussian-to-mesh baselines take tens to hundreds of seconds. Trained on RealEstate10K and DL3DV and tested zero-shot on ScanNet, it’s aimed squarely at robotics — locomotion, dynamics, and grasping all need scenes you can simulate, not just look at. Turning a phone-style image sequence into a usable physics scene in a second moves real-world 3D from offline asset prep toward something closer to real-time.

Leave a comment