A splatting method rendered captured scenes in real time

The process wall

Photogrammetry and neural radiance fields could already turn a set of photographs into a navigable scene by 2023, but rendering a finished capture smoothly, at high resolution, in real time, still cost more than most working pipelines could spend per frame. The 3D Gaussian Splatting paper, submitted to arXiv on 8 August 2023 and published in the SIGGRAPH 2023 issue of ACM Transactions on Graphics, states plainly that for 'unbounded and complete scenes... at 1080p resolution, no current method can achieve real-time display rates.' Its method starts from the same sparse point cloud a camera-calibration step already produces, and represents the scene as a set of 3D Gaussians rather than a neural network queried per pixel.

What the documents show

The paper describes three components: representing the scene as anisotropic 3D Gaussians derived from calibration points, an interleaved optimisation and density-control step that adds or removes Gaussians as training proceeds, and a visibility-aware rasteriser built to splat those Gaussians directly rather than march rays through a volume. The arXiv abstract states the result reaches 'real-time (>= 30 fps) novel-view synthesis at 1080p,' while the project page hosted by Inria states a higher figure, 'greater than or equal to 100 fps,' for the same claim. That gap between two documents from the same team is worth noting rather than resolving by preference.

Authorship and limitation

The mechanism, splatting explicit, optimisable primitives instead of querying an implicit network, is what removes the render-time cost a NeRF-style method pays per pixel, at the cost of a representation made of millions of individual Gaussians rather than one compact network. That representation is not automatically free to reuse: the project's own licence states that Inria and the Max Planck Institute for Informatics 'hold all the ownership rights' on the software, grants use for research, testing and evaluation, and requires contacting Inria directly for any other use, including anything commercial.

What survives

What a production keeps from a Gaussian-splat capture is the same asset a NeRF capture would have kept, the source photographs and their recovered camera poses, since the Gaussians themselves are a derived, re-optimisable fit to that data rather than hand-authored geometry. Editorially, real-time captured scenes of this kind are most immediately useful for set scouting, previsualisation and archiving a location as it existed at capture time, applications the paper's speed claim supports even though the paper itself does not discuss virtual production use.

• Check which licence governs a specific Gaussian-splatting implementation before any commercial use, since the original code is research-only.
• Compare a demonstrated frame rate against the resolution and scene complexity it was measured at, given the paper's own two stated figures.
• Ask whether a captured scene retains its original camera poses and source imagery, or only the trained Gaussian representation.

Read together, the paper and its licence describe a specific trade, rendering speed bought by representation and optimisation choices, released for study but not, on the authors' own terms, for unrestricted reuse.