Skip to content
Technology / Benchmarks

Don’t trust a number. Orbit the real thing.

Instead of a leaderboard, this page gives you our own measurements on real reconstructions: how long a scene takes to solve inside the app, how fast it renders afterwards, and how small it gets on the way out. Every figure is our own instrumentation — reproducible, dated, and stated without comparison to anyone else. Orbit an actual reconstruction yourself on the Live Viewer.

The real thing

Start by judging a reconstruction, not a number

No video, no pre-baked turntable: this is one of our own reconstructions decoding in your browser as moving 4D Gaussians. Drag to look around and scrub the take. Everything measured further down this page was produced by the same pipeline that made it.

Scene: ViVo “athlete” (Azzarelli et al., 2025, University of Bristol), licensed CC BY 4.0 and reconstructed as 4D Gaussians by VoxaVerse. It is the only scene this site hosts. The research-licensed datasets cited below appear as measurements and still frames — never as scenes you can stream, bookmark or download. Open it full-screen on the Live Viewer, where you can also open a scene of your own.

Solve time

How long a scene takes to make, inside the app

Not an engine microbenchmark. Each of these is one run of the Mac studio driven through its own interface: import the capture the way a person does, click Solve, and stop the clock when the capture’s own card says it finished. Preparation, motion initialisation and the solve are all inside the number, because all three are inside the wait.

SceneSolve timeWall clock
flame_steak1 h 27 m5,224 s
cut_roasted_beef1 h 39 m5,933 s
cook_spinach1 h 40 m5,980 s
coffee_martini1 h 57 m7,041 s
flame_salmon_11 h 59 m7,149 s

Master preset · 30,000 iterations · 1600 px working resolution · subject isolation off. Measured on Apple M3 Pro, 36 GB unified memory, captured 2026-08-26, on the shipping build with nothing else contending for the GPU. A sixth scene of the same set finished on the same run, but the harness timing it lost the reading — so it is missing from this table rather than estimated into it. No quality score is published beside these times: that run did not record a final evaluation figure per scene, and a number nobody can point at does not belong on this page.

Evaluation still from one of the timed runs: the cut_roasted_beef scene part-way through solving in the studio viewport, with a telemetry rail beside it reporting the loss curve, the evaluation curve, the splat count, the iteration rate and the memory in use
Evaluation still — cut_roasted_beef part-way through the run timed above, with the run reporting on itself beside it. Neural 3D Video dataset, CC BY-NC 4.0; shown at figure size, not hosted.

Scenes named above and shown here belong to the Neural 3D Video dataset (Li et al., Neural 3D Video Synthesis from Multi-view Video, CVPR 2022, Meta AI), under CC BY-NC 4.0. Cited as measurements and shown as still frames only — not hosted, not streamed, not offered for download. Source footage and model rights remain with their owners.

Render performance

Interactive frame times, measured on-device

400-frame render benchmarks on the engine’s own native Metal player, pruned production checkpoints — our instrumentation, no synthetic best case, no comparison to anyone else.

Scene named below: the Neural 3D Video dataset (Li et al., Neural 3D Video Synthesis from Multi-view Video, CVPR 2022, Meta AI), distributed under CC BY-NC 4.0. Cited here as a measured number, not hosted or streamed — source footage and model rights remain with their owners.

cook_spinach · 1352×1014
6.61 ms
mean GPU frame (~151 FPS) · p99 7.30 ms
Orbit ceiling (native player)
800+ FPS
Metal, pruned checkpoint, small scenes

Measured on Apple M3 Pro, 36 GB unified memory, captured 2026-07-08. Browser and headset delivery through Vortex targets a different budget — see /technology/standards for the streaming story.

Delivery size

Compression across every open interchange format

One representative 345,892-splat, 5-second scene, exported to every format the engine supports today — measured from the export tool, not modeled.

FormatSizeKeeps time?Primary consumer
.ply37.3 MBNo (single moment)Universal — every splat tool
.usdz20.3 MBNo (single moment)RealityKit / visionOS / Quick Look / Omniverse
.spz3.9 MBNo (single moment)Scaniverse, SPZ-compatible loaders
.sog2.5 MBNo (single moment)SuperSplat, PlayCanvas
.nx4d36.1 MB — the full 5 s of 4DYesVortex — the only format that keeps time

The .nx4d figure carries the entire 5-second free-viewpoint temporal scene — about 7.2 MB per second of content — without per-frame keyframing. Measured 2026-07-08from the engine’s export tool.

Why there’s no leaderboard here

We don’t publish quality-score comparisons against other systems. A fair comparison needs identical scenes, identical metrics, identical protocols, and independently reproducible outputs — and until we can run one to that standard, a number would be marketing, not evidence. So instead of a leaderboard, this page gives you our own reproducible solve, render and size measurements, plus a real reconstruction you can orbit for yourself at the top of it. Nothing on this page is a claim about anyone else’s product.

See it move — not just the numbers.

The reconstruction at the top of this page is streaming live. The Live Viewer gives it a full screen, and room for a scene of your own.