Everyone’s framing splats as LED wall content. For a studio our size the interesting part isn’t how you display them — it’s that they change the arithmetic of building an environment at all. Here’s where we think they earn their place, and where they don’t.
A client rang us a few weeks back wanting their own workshop as the backdrop for a run of product films. Lovely space — timber, dusty light, a lot of character. The honest answer was that modelling it properly in Unreal is a two-to-four week job, which is more build time than the shoot itself justified. That conversation happens more often than you’d think, and it’s usually where a good idea quietly dies on cost.
Gaussian splatting is the thing that’s been steadily rearranging that maths. Volinga added 4D splat playback to its Unreal pipeline at the start of this month, which is the update that pushed us to write this — but the wider story has been building for a year and a half, and almost all of it has been told from an LED volume point of view. We think that framing undersells it for studios like ours.
What a splat is, in plain terms
A 3D Gaussian splat isn’t a model. There’s no mesh, no UV map, no PBR texture set. You photograph or film a real place from a lot of angles, and a training process reconstructs it as millions of tiny, oriented, semi-transparent blobs — each one carrying a position, a size, an orientation, and a colour that changes depending on where you view it from. Stack enough of them and you get something that reads as photographic from any angle inside the captured space.
That last bit is the whole trick. A 2D plate is photographic but flat. A built 3D environment has real geometry but takes weeks and never quite loses that “someone made this” quality on close inspection. A splat is photographic and has parallax, and you can capture one in an afternoon.
For us the practical route in is capture, process, then drop the result into Unreal as an asset alongside everything else. Volinga’s suite encodes the capture into its own NVOL format and its plugin brings it into the content browser like any other object, with Sequencer and sublevel streaming now in the mix. That’s the bit that matters — not the novelty of the technique, but that it behaves like a normal set piece in a scene we already know how to light and comp.
Why we think this is a green-screen story too
Read the coverage and you’d assume splats exist to be shown on an LED wall. Netflix put Calle Alcalá on a volume that way. Pixotope integrated Volinga for broadcast. Every case study is a wall.
But the wall is the display. Splats aren’t a display technology — they’re an acquisition technology, and acquisition is the expensive half of virtual production for everyone below the top tier. On our green-screen stage at DocShed the environment is rendered in Unreal and composited live in Assimilate Live FX against the keyed Blackmagic Pyxis 6K feed. It makes no difference to that chain whether the environment is a hand-built set or a splat of a real building — but it makes an enormous difference to the schedule and the quote.
There’s a second advantage that’s specific to compositing rather than in-camera work. A splat carries the real lighting of the place it was captured in, baked in. On an LED volume that’s a mixed blessing, because the wall is also your light source and baked lighting limits what you can do with it. On green screen, the background lighting is reference, not illumination — we read the direction, colour and contrast of the capture and then build the physical key, fill and wrap to match it, which is the job we’ve written about before. Baked lighting we can measure is genuinely easier to match than a CG environment where we chose the lighting ourselves and have to keep two mental models in sync.
Where they fall over, honestly
We’re not about to pretend this is free. Four things bite.
The capture envelope is the shot list. A splat only holds up inside the volume of space you photographed it from. Push the virtual camera outside that envelope and it degrades fast — smeary, translucent, wrong. That means the capture has to be planned against the intended camera moves, not just “walk round with a camera for twenty minutes and see”. In practice it turns a location recce into a technical exercise, and if the director changes their mind about a move on the day, you can’t always give it to them.
Relighting is still the weak point. Baked lighting is helpful for matching, as above, but it’s fixed. If a scene needs to move from afternoon to dusk, a built Unreal environment does that in a slider and a splat does not. Relighting tools are improving, but they’re not at the point where we’d promise a client a time-of-day change on a splat.
There’s no geometry to interact with. No collision, no shadow catching, no physics, and editing is closer to sculpting a point cloud than moving objects around. If the brief involves a virtual object that has to sit in the world rather than behind talent, you’re back to building.
Grazing angles and thin detail. Foliage, wire, chain-link and fine edges are where the artefacts live. Front-on it’s astonishing; at a shallow angle across a detailed surface it can go to mush.
So our position is this: a splat is not a replacement for a built environment. It’s a genuine third option sitting between a 2D plate and a full 3D build — more depth and camera freedom than a plate, a fraction of the build time of a bespoke set, and a narrower range of things you’re allowed to ask of it. Knowing which of those three a project actually needs is most of the job, which is roughly the argument our VP Decision Maker exists to make.
Why the 4D update matters more than it sounds
The thing that has always given static splats away behind a performer is that they’re dead. Real places move. Leaves shift, water moves, a curtain breathes, someone walks past a window. A frozen background sat behind a living actor sets off the same alarm in a viewer’s head as a bad key does — they can’t name it, but they know.
4D splatting extends the reconstruction into time, so a capture can hold motion rather than a single frozen moment. Volinga’s implementation collapses what would otherwise be a per-frame stack of point clouds into one encoded asset the engine can stream, which is the practical hurdle rather than the theoretical one. That’s the difference between a technique that demos well and one you can hang a shoot day on.
It also puts a timing question straight back on our desk. Anything with motion in it has to sit on the same beat as everything else in the chain — which is exactly the genlock and frame sync problem we wrote about a month ago, now with one more clock in the room.
What’s next
We’re testing capture methodology through the back end of this year, mostly around North Wales — the coast, some interiors, a couple of industrial spaces — with a specific focus on how tightly we can plan a capture against a storyboarded camera move rather than capturing broadly and hoping. Our instinct is that the discipline is in the recce, not the software.
We’re also curious whether a splat and a built set can share a frame cleanly: a photographic mid-ground from a real location with hand-built geometry in front of it for anything talent needs to interact with. On green screen, where we’re compositing layers anyway, that hybrid looks more achievable than it would in camera on a volume. We’d rather find the failure modes on our own time than on a client’s shoot day — which is the same reason we ran five virtual environments across a single day for Aber Falls only after we knew the switching held up.
If you’ve got a real location you’d love to shoot in but can’t — because it’s busy, expensive, three hours away, or simply no longer looks like it did — that’s exactly the conversation worth having. Book a studio visit and we’ll show you what a capture looks like on the monitor next to a built set, and you can judge the difference with your own eyes.
This Fortnight in VP
Volinga adds 4D Gaussian splatting to its Unreal Engine pipeline — Animated splat captures now play back in real time inside Unreal as a single encoded asset, with UE 5.8 support, Sequencer integration and sublevel streaming alongside. Motion is the thing static splats couldn’t do, and it’s what makes them usable behind a performer.
Brompton processing powers the LED volume behind ITV’s Believe Me — fivefold studios recreated nighttime London cab journeys on a 30m × 4m ROE Visual wall, and reckon the same scenes shot on location would have taken three to five times longer. A useful reminder that VP’s strongest argument is still schedule, not spectacle.
Tencent’s WorldClaw paper describes agentic text-to-3D that returns editable assets — Generated scenes come back as instance-level objects rather than one locked render. Editability is the entire difference between a nice image and something a production can actually build a set from.
LTX-2.5 launches with multi-shot continuity, native 4K HDR and a RAW finishing workflow — Generated shots that hold character, environment and lighting across cuts, output into a proper colour pipeline. Whatever you make of the ethics, the pipeline end of it is getting harder to dismiss.
The ASC is open-sourcing StEM3-VP, a virtual production test kit for LED volumes — Open reference assets for evaluating and calibrating ICVFX stages, hosted in the Academy Software Foundation’s library. A shared, vendor-neutral yardstick is exactly what smaller stages need to prove themselves without a week of guesswork.
