Orbit the moment. While it plays.
VoxaVerse Prism is a native Mac studio for volumetric scenes. Capture with a rig of iPhones or a set of photos, Solve on your own Mac, and then move the camera anywhere through a recorded moment as it runs — not a video of a viewpoint someone else chose.
Works standalone. Prism opens and plays every major splat format on its own, and Solves from captures you already have. It does not require a Beacon rig, an account, or the rest of the pipeline.
Capture. Import. Solve. Share.
Four moves, one window, no service in the middle of any of them.
Capture
A rig of iPhones, or the photos you already have.
Set VoxaVerse Beacon running on two or more iPhones and they record the same seconds from every side, on one shared rig clock, with LiDAR depth alongside the colour. No rig? A photo set works: bring a COLMAP or RTAB-Map export and Prism builds a still scene from it.
See the capture rigImport
Drop the session in. Any rig, converted once.
A Beacon session imports as itself. Anything else arrives through the open Capture Package format, which states its own axes, units, depth encoding and timing instead of leaving a reader to guess — so a dataset from someone else’s rig lands with the same certainty as one from your own.
The open capture formatSolve
Watch the scene resolve, on your machine, in the viewport.
Solving is not a progress bar over a black box: the viewport shows the actual model as it converges, and you can orbit it while it does. Pause, continue or stop without losing a step — Prism says exactly how far a scene has come and where Continue picks up. Every control maps to a real engine parameter, and Auto means the engine’s own judgement rather than a hidden preset.
Share
Open formats out — including one that keeps time.
Export to the formats the rest of the industry reads, or to a VoxaVerse Scene, which carries the whole moving scene in a single file. Drop that file into the viewer on this site and it opens in the browser, on your machine, with nothing uploaded.
Open a scene in the viewer
Every control is a real control.
The inspector is not a wrapper around three presets. Schedule, densification, the coarse-to-fine pyramid, the loss terms — each maps to a parameter the engine genuinely uses, and each says what the engine would choose on its own.
- Fast, Medium or HD to start, then open any of it up. Auto is the engine’s judgement, shown as a value rather than hidden.
- Pause, continue, stop — none of them lose a step, and the app states exactly where Continue resumes.
- Region cuts while it Solves, with undo that restores the exact model from before the cut.
- Held-out views so the scene is scored on frames it was never shown.
An evening, on a laptop.
These are runs of this app, on one MacBook Pro, driven through its own interface — click Solve, and stop the clock when the capture's card says it finished. Preparation and motion initialisation are inside the number, because they are inside the wait.
| Scene | Solve time | Wall clock |
|---|---|---|
| flame_steak | 1 h 27 m | 5,224 s |
| cut_roasted_beef | 1 h 39 m | 5,933 s |
| cook_spinach | 1 h 40 m | 5,980 s |
| coffee_martini | 1 h 57 m | 7,041 s |
| flame_salmon_1 | 1 h 59 m | 7,149 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. A sixth scene of the same set finished on the same run, but the harness timing it lost the reading, so it is missing here rather than estimated. Scenes belong to the Neural 3D Video dataset under CC BY-NC 4.0 — cited as measurements and shown as still frames only, never hosted or streamed. The full methodology is on the benchmarks page.
What you get out, and who can read it.
A scene you cannot take anywhere is not an asset. Prism writes the formats the rest of the industry already reads — and reads its own exports back, so a round trip is something you can check rather than trust.
| Export | What it is | Who reads it |
|---|---|---|
| PLY | The reference Gaussian-splatting binary layout | Blender add-ons, SuperSplat, effectively every splat tool |
| SPZ | Niantic’s compact splat container | Scaniverse and SPZ-compatible loaders |
| SOG | Compact web splats | SuperSplat, PlayCanvas, browser viewers |
| glTF / GLB | Khronos KHR_gaussian_splatting | Passes the Khronos validator with zero errors and zero warnings |
| USDZ | Time-sampled OpenUSD with the Gaussian-splat schema, plus a preview mesh | Quick Look shows it today; splat-aware renderers get the full field |
| VoxaVerse Scene | The native volumetric container — the only one here that keeps time | The Vortex web viewer, on this site, in any browser |
Prism also writes numbered frame sequences for tools that would rather have one file per moment. Its round-trip check puts the live scene beside its own re-imported export with a single camera driving both, so “it exported fine” is something you can see rather than assume.
Offline isn’t a mode. It’s the architecture.
The volumetric category runs on upload-and-wait. Prism doesn't, and the reason is structural rather than a policy that could quietly change.
Solving runs on your Mac, on your footage. There is no upload step, no queue and no per-scene compute bill — the work happens where the files already are.
Captures and scenes live on your own disk, in your own storage, browsable from the app. No account is required to open the app or to use it.
Optional subject isolation is an on-device model. The frames it looks at never leave the machine, and it can be switched off entirely.
Pull the network cable and the studio behaves identically. Nothing about the pipeline needs a connection, which is also what makes it usable on a set with no reliable one.
Native, and it scales.
Prism is built for Apple silicon end to end — Metal for both solving and playback, with no Python runtime, no translation layer for someone else's GPU architecture, and no web view pretending to be an app.
That is what lets one build behave sensibly across the whole Mac range. On an M-series MacBook it works within the memory and thermal envelope a laptop actually has. On an Ultra-class desktop — up to and including the designs that carry two Neural Engines — it uses the wider memory bandwidth, the larger GPU and the extra on-device inference capacity it finds, rather than asking you to pick a build or a tier.
Quality tiers are measured on your machine rather than assumed from a model name, and the app tells you what it measured.
The parts that only matter when they're missing.
PLY, SPZ, SOG, glTF, USDZ and VoxaVerse Scenes — including files that came from tools with no connection to us.
Orbit, pan, dolly, roll, frame and camera presets, on the bindings 3D artists already have in their hands.
Play, scrub, loop and frame-step a moving scene at its own true pace.
Including the 3D viewport. Reduce Motion and Reduce Transparency are honoured, not ignored.
Viewing is free, and stays free. Creating scenes unlocks with a one-time purchase or a subscription after a free trial — and whatever you choose, everything you have already made stays yours and stays openable. Building on the open Capture Package is part of the same commitment: the specification and its validator are public.
The scenes on this site came out of it.
Every volumetric scene you can orbit on voxaverse.ai was solved with this engine and exported from this pipeline. Leave an address and be first through the door when the studio ships.