Reshooting a Scene After Filming: Gracia Wants to Reconstruct Reality in 4D

Gracia unveils a 4D reconstruction model to recreate filmed scenes from new angles, with applications in cinema, sports, and robotics.

Changing a camera angle after a shoot, observing an athletic movement from a perspective that was never filmed, or integrating an actor's real-life performance into a virtual environment. These possibilities are at the heart of Gracia's World Reconstruction Model, an artificial intelligence system capable of transforming multi-camera recordings into animated three-dimensional scenes. Its key distinction: prioritizing the geometric reconstruction of reality over generating images from a free interpretation of events. Reconstructing a scene without inventing what happened

Generative video models can create new viewing angles, but they also risk inventing elements that were absent from the recordings: details of an object, the position of a limb, or a character's movement.

Gracia aims to reduce this uncertainty by reconstructing geometry and motion from actual footage. The result takes the form of a 4D representation, combining a three-dimensional environment with its evolution over time.

The system relies notably on 4D Gaussian Splatting (4DGS), a technique that allows animated volumetric scenes to be represented and rendered from different angles.

Gracia claims sub-millimeter geometric accuracy and faithful rendering relative to the recordings. However, at this stage, this accuracy is not accompanied by a public protocol allowing the conditions under which it is achieved to be verified. The company also presents its approach as hallucination-free reconstruction, a claim to be distinguished from an absolute guarantee of accuracy. Four to eight cameras to replay sports action

One of the most concrete applications involves sports broadcasting. Instead of relying solely on the angles filmed during an action, directors can reconstruct the scene and then virtually move the camera around the participants.

In its sports presentation, Gracia indicates that its new system can operate with four to eight cameras and produce slow-motion replays from free-view angles in near real time.

By comparison, traditional volumetric reconstruction methods often require more than 40 specialized cameras and at least 24 hours of processing time.

Gracia's technology was previously used at the 2026 PGA Championship, in collaboration with Radiant Images. For that production, the movements of several golfers were captured using 56 iPhone 17 Pros, before being reconstructed for television broadcast roughly 24 hours later.

This demonstration relied on an earlier 4DGS setup and should therefore not be confused with a validation of the new four-camera system and its near-instantaneous processing. Actors become volumetric performances

In cinema and visual effects, the benefit lies in preserving a genuinely filmed performance while recovering camera framing flexibility during post-production.

An actor captured in 4D can be integrated into a digital scene, observed from another angle, or paired with a virtual environment. The director thus retains an existing performance without necessarily having to recreate their movements through animation.

Gracia notably cites the music video Nightshift Superstar by Muse, directed with Radiant Images, in which Matt Bellamy was turned into an animated volumetric representation.

The Gracia VFX workflow also includes possibilities for lighting modifications and integration into tools like Unreal Engine and Unity.

Nevertheless, this freedom depends on the quality and coverage of the initial shots. Even an advanced reconstruction cannot guarantee the presence of details that no camera properly captured. Physical environments to train robots

Gracia is also targeting robotics with so-called Real2Sim2Real applications. The principle consists of capturing real-world situations, reconstructing them in a digital environment, and then using this data to train or evaluate robotic systems before returning them to the physical world.

The technology can notably provide information on depth, object positions, and hand movements. This data can help test behaviors or compare a robot's performance across reproduced situations.

The challenge remains the physical fidelity of the reconstructed environments. A visually realistic scene is not always sufficient to accurately describe the contacts, forces, or material properties necessary for certain robotic simulations. From the set to the 3D engine

Gracia presents its system as camera-agnostic, even though capture setups must meet technical constraints, particularly the hardware synchronization of shots.

Its tools include a 4DGS processing suite and SDKs designed for Unity, Unreal Engine, the web, Python, C/C++, and Apple platforms. The company also offers streaming solutions for web browsers and immersive headsets.

Services remain primarily structured around professional projects. On its standard 4DGS offering, Gracia lists processing rates starting at $800 per minute for streaming and virtual reality, and $2,500 per minute for VFX uses. These prices do not necessarily reflect those of the new World Reconstruction Model.

With this technology, Gracia is exploring a different direction from that of traditional video generators. It is no longer a matter of asking an AI to imagine a scene, but rather having it reconstruct an event that was already captured so it can be examined, modified, or reused.

This distinction is essential for use cases where facts matter as much as the image, whether it involves verifying an athletic movement, preserving an actor's performance, or building robotics research data. The question going forward will be measuring how far this fidelity can be maintained with fewer cameras and reduced processing times.