Every number here is dated, sourced, and re-runnable.
No one else in this category publishes a benchmarks page that pairs reconstruction quality with transparent methodology. This page does — and it flags anything not yet independently reproduced instead of presenting it as fact.
Don't take the table's word for it — orbit the benchmark itself
These are the actual reconstructions the numbers below were measured on: real Neural 3D Video dataset scenes, reconstructed by the VoxaVerse engine, streaming as native 4D Gaussian scenes. Switch scenes, scrub time, drag to look around.
Dataset attribution & licensing
The evaluation scenes above are reconstructed from the Neural 3D Video dataset (Li et al., Neural 3D Video Synthesis from Multi-view Video, CVPR 2022, Meta AI), distributed under the CC BY-NC 4.0 license. The juggle row in the table below cites the CMU Panoptic Studio dataset (Joo et al., ICCV 2015), and the demo scenes on the home page are synthetic sequences from the D-NeRF dataset (Pumarola et al., D-NeRF: Neural Radiance Fields for Dynamic Scenes, CVPR 2021). All of these appear as non-commercial research demonstrations of reconstruction quality — they are benchmarks, not products, and are credited wherever shown. Source footage and model rights remain with their respective owners.
PSNR against published baselines, per scene
Held-out camera protocol — the training loop never sees these views. Splat counts and scene identities are stated; nothing is cherry-picked to a single number.
| Metric | Value | Status |
|---|---|---|
| juggle (Panoptic Sports, 27-cam 360° dome, 688k splats) | 30.37 dB vs published Dynamic3DGaussians baseline: 29.48 dB — official 4-camera held-out protocol, 150 frames. | Reproduced |
| cook_spinach (N3V/DyNeRF, two-stage schedule) | 32.49 dB (masked, crisp) 30k backbone → prune → 7k masked fine-tune, ~2h25m end-to-end on the hardware below. | Reproduced |
| cook_spinach — held-out cam00 | 33.10 dB | Reproduced |
| flame_steak — held-out cam00 | 33.83 dB | Reproduced |
| coffee_martini / flame_salmon_1 | in flight — two hypotheses tested and refuted Seed-cloud and camera-desync hypotheses were both measured and ruled out. Root cause still open — published honestly rather than omitted. | Internal — pending validation |
Interactive frame times, measured on-device
400-frame render benchmarks on pruned production checkpoints — no GPU-timeline overflow, no synthetic best case.
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.
Compression across every open interchange format
Same checkpoint (juggle, 345,892 splats), exported to every format the engine supports today — measured, not modeled.
| Format | Size | Keeps time? | Primary consumer |
|---|---|---|---|
| .ply | 37.3 MB | No (single moment) | Universal — every splat tool |
| .usdz | 20.3 MB | No (single moment) | RealityKit / visionOS / Quick Look / Omniverse |
| .spz | 3.9 MB | No (single moment) | Scaniverse, SPZ-compatible loaders |
| .sog | 2.5 MB | No (single moment) | SuperSplat, PlayCanvas |
| .nx4d | 36.1 MB — the full 5 s of 4D | Yes | Vortex — the only format that keeps time |
The .nx4d figure carries the entire 5-second free-viewpoint temporal scene — 7.2 MB/s of content, inside 4DV.ai’s own published delivery band (3.75–7.5 MB/s) — without per-frame keyframing. Measured 2026-07-08from the engine’s export tool.
The vs-Luma claim, and why it isn't a number yet
Internal — pending third-party validation
The VoxaVerse engine’s no-COLMAP reconstruction is engineered to compete directly with Luma AI at static reconstruction quality. As of 2026-07-08, we have not yet runa matched, apples-to-apples comparison against Luma on identical scenes with identical metrics — so no number is published here. When that comparison runs, it will appear on this page with the exact scenes, the metrics (PSNR/SSIM/LPIPS), the capture date, and the raw outputs — the same standard every other row on this page is held to. Until then, treat any “beats Luma” language anywhere as a stated engineering target, not a measured result.
See it move — not just the numbers.
The N3V scenes benchmarked above are streaming at the top of this page — and the Live Viewer has more. Orbit them yourself.