Dev recreates Tesla Reality View using splats
Developer Ryan Vogel is reverse-engineering Tesla's FSD Reality View by mapping Model 3 camera arrays and generating temporal 3D Gaussian splats of road surfaces. The build reconstructs road geometry and depth from multi-view driving video to benchmark against Tesla's proprietary perception stack.
Reconstructing Tesla's proprietary perception stack from consumer vehicle feeds highlights how mature multi-camera neural rendering and 3D Gaussian splatting have become for individual developers.
- –Temporal Gaussian splatting from sequential driving video creates continuous metric depth representations without relying on expensive active LiDAR hardware.
- –Mapping Model 3 camera intrinsics and extrinsics enables accurate multi-view synchronization across vehicle angles into a unified coordinate system.
- –Independent reverse-engineering efforts help demystify the closed, end-to-end vision architectures powering commercial autonomous driving systems.
- –The primary hurdle remains scaling live neural rendering from static road surfaces to dynamic multi-agent surroundings on constrained edge hardware.
DISCOVERED
10m ago
2026-11-21
PUBLISHED
21m ago
2026-11-21
RELEVANCE
AUTHOR
ryanvogel