Gracia has publicly introduced its World Reconstruction Model, which turns real multi-camera footage into an accurate 4D representation of a scene.
Gracia built the model for Physical AI and for applications where hallucinations are not acceptable. The company names four uses, free-viewpoint replays in near real time for sports broadcasting, real performances in full detail for VFX and film, a self-service reconstruction suite for 4D Gaussian splatting (4DGS), and Real2Sim2Real, evals and benchmarks for robotics. Gracia says ground-truth detail and sub-millimeter geometric accuracy carry across all four. It has not published how that accuracy is measured.
Gracia's broadcasting page sets the model against two alternatives. Classic 4D reconstruction, by Gracia's figures, needs 40 or more specialized cameras and a minimum 24-hour turnaround. The World Reconstruction Model needs 4 to 8 cameras and runs in near real time, while single-camera AI video generation, Gracia says, hallucinates geometry and details. When Radiant Images captured four golfers at the 2026 PGA Championship on 56 iPhone 17 Pros, Gracia's 4DGS turnaround was about 24 hours from receipt of footage.
Gracia describes the World Reconstruction Model as hardware-agnostic, with footage from real-world cameras going in and a 4D scene coming out. The FAQ on its 4DGS page narrows that. Capture requires shutter sync and genlock across hardware-synchronized cameras, with iPhone 17, Blackmagic and Sony FR7 named as options and GoPro and Insta360 ruled out for unreliable sync. Gracia handles processing only and leaves capture to partner studios.
For rendering, Gracia's announcement lists integrations with Unity, Unreal Engine, Web SDK, Python, C/C++, Apple platforms and streaming infrastructure. The Web SDK plugs into React, Three.js and PlayCanvas, the Python SDK renders 3DGS and 4DGS frames to NumPy with color and depth, and the C/C++ SDK pairs a C API with a header-only C++ wrapper on Vulkan. The Apple SDK is a Metal renderer for macOS, iOS and visionOS. Gracia's codec now carries 200K splats per frame at 75 Mbps, against the 120,000 splats per frame its Apple Vision Pro streams used at the same bitrate in June.
For robotics data, Gracia lists metric depth, 21-joint 4D hand tracking and per-splat semantic attributes.
More on the World Reconstruction Model is available on Gracia's site.
