4D Gaussian Splatting Is Arriving on Set. Genlock Still Matters.

There is a version of the 4D Gaussian splatting story that gets told at trade shows, and there is the version you learn standing in a stage bay at 6am with a first AD asking how long the setup is going to take. Both versions are true. They are just answering different questions.

The trade-show version had a strong year. At SIGGRAPH’s Real-Time Live!, InfiniteStudio showed the first 4D volumetric capture system genuinely aimed at professional video production rather than research demos — production-level quality, long-duration takes, and compression landing in the 100–200 Mbps range. At NAB, 4DV.ai and OBSBOT put up a sixty-camera rig built entirely from PTZ cameras. 4DViews moved its pipeline onto Gaussian splatting as well. On the research side, Detail Enhanced Gaussian Splatting pushed large-scale volumetric capture quality up another step, and OTOY has said Octane will render and relight splats directly — which is the part that turns a novelty into a deliverable.

We have been building toward this for a while. Our modular volumetric rig, C.A.L.V.I.N.N., exists because we started asking these questions in 2017 on Watchmen, long before anyone called it 4DGS.

What the new systems actually solve

The genuine breakthrough is not resolution. It is temporal continuity.

Older volumetric pipelines reconstructed each frame more or less independently, then stitched the results into a sequence. That is why the classic volumetric artifact is a kind of boiling — geometry that swims and reshuffles between frames even when the performer is barely moving. 4DGS treats the splats as a continuous representation through time rather than a stack of per-frame solves. Track a splat across the take instead of re-deriving it, and the boiling largely goes away.

Two things fall out of that, and both are worth a producer’s attention:

Slow motion becomes free. If the representation is continuous in time, you can sample it at any frame rate. You are no longer capped by what the cameras shot. That is a real change to how a stunt or a hero beat can be covered.

Camera decisions move to post. This is the bigger one. A conventional shoot commits to position, lens, and angle at the moment of capture. A well-captured 4D asset lets editorial re-block the coverage weeks later. That is the same argument that sold everyone on virtual production, arriving now for performance rather than environment.

What they quietly don’t solve

Here is the claim in the 4DV.ai coverage that we want to be careful about, because it is going to get repeated: the tolerance for unsynchronized cameras.

It is a real result and it is genuinely useful. Tolerance is not the same as indifference, though. What that tolerance actually buys you is robustness — a rig that degrades gracefully instead of failing when sync drifts, which matters enormously for a sixty-camera PTZ array assembled fast. It is a very good property to have.

What it does not buy you is a clean plate on a show with a VFX supervisor who needs the volumetric asset to sit in the same frame as principal photography, matched to a real camera, at 23.976 with the rest of the production. The moment your capture has to integrate rather than stand alone, sub-frame offsets stop being a reconstruction problem and start being a compositing problem. You can solve boiling in the reconstruction. You cannot solve a performance that is eleven milliseconds out of step with the plate it has to live in.

So we genlock. Every camera on our arrays, every time. Not because the new solvers are weak — they are impressively strong — but because sync is the cheapest thing to get right on the day and one of the most expensive to fix in post. The tolerance is a safety net. We would rather not need one.

The other thing the demos tend to skip: a 4D asset at 100–200 Mbps is a lot of data per second of screen time, and it has to move through a facility that was built around image sequences and caches. Our Mobile CyberSCAN™ Studios are configured to handle that on location precisely because pulling this volume of data back to a facility before anyone can look at it is how a schedule gets eaten.

Where we think this lands

Our read, plainly: 4DGS is not going to replace scan-and-rebuild digital doubles in the next two years, and anyone telling a producer otherwise is selling something. What it is going to do is take over the middle of the market — the crowd performer, the background beat, the stunt pass, the shot where you need a real human at real fidelity for four seconds and cannot justify a full digital double build.

That is a large and currently under-served part of most VFX budgets. It is also where the technology is strongest right now, because a four-second continuous take is exactly what these representations handle best.

What we would tell a supervisor scoping this today: capture with sync even if your solver says it does not need it, plan for the data rate before the shoot rather than after, and pick your shot for the technology’s strength — continuous performance, moderate duration, no extreme close-up — rather than trying to make it do everything. The 4D pipeline is production-ready for the right shot. Picking the right shot is still the job.

If you have a show where this fits, tell us what the shot needs and we will scope it honestly — including telling you when a conventional capture is the better answer.