Fast 3D Foundation Model Initialized Gaussian Splatting

A new reconstruction pipeline pushes 3D Gaussian Splatting closer to practical production use by removing the usual Structure-from-Motion bottleneck. Instead of starting from a carefully built sparse reconstruction, it initializes camera poses and point clouds from 3D foundation models, then jointly optimizes the poses and Gaussian primitives with depth-guided supervision.

The big draw is speed. The method is designed to converge from rough initialization with only 50-60 input views, and the reported training time is around three minutes per scene. That makes it much easier to imagine rapid capture workflows where a scene can be turned into a ready-to-use 3DGS asset without a long preprocessing stage.

Quality is still competitive rather than purely speed-first. On Mip-NeRF 360, Tanks and Temples, and RobustNeRF, the method reports 23.61 dB PSNR and 0.19 LPIPS, while also adding an MLP-based pose refinement module to help sparse-view cases. For graphics teams, the practical takeaway is that foundation models are starting to do more of the heavy lifting in reconstruction pipelines, especially when input coverage is limited.

The broader implication is that 3DGS workflows may become much more accessible for near-real-time applications. That matters for robotics, VR, and autonomous navigation, but it also points toward faster scene ingestion for game-adjacent tools, virtual production, and any pipeline where capture-to-preview time is a constraint.