MetaLoom // Lumen · spatial media engine

Video goes in.
A place comes back.

Lumen turns ordinary footage into navigable 3D Gaussian-splat scenes, stores them the way a database stores records, and makes physical places searchable — from a photo, or from a sentence.

Talk to us about Lumen How it works

Commercial software — a container image for your own GPU hardware

Scroll

Gaussian splat generation

From flat frames
to a field of light.

Lumen recovers where the camera stood in every frame, then optimises millions of anisotropic 3D Gaussians — position, shape, colour, opacity — until rendering them reproduces the footage. What comes out is not a video of a place. It is the place, viewable from anywhere.

How a video becomes a Gaussian-splat scene Video frames on the left; structure-from-motion recovers the camera pose of each frame around a sparse point cloud; a differentiable rasteriser then optimises millions of anisotropic 3D Gaussians until rendering them reproduces the footage. video frames poses recovered millions of Gaussians render ↔ compare ↔ adjust
held-out PSNR 26.70 dB reference implementation: 27.41 dB ~25 min on one GPU
  1. 01

    Recover the cameras

    Structure-from-motion reads the parallax between frames and works out the pose of every shot — no rig metadata, no markers, just the video.

  2. 02

    Optimise the scene

    A differentiable rasteriser renders the current guess, compares it against the real frames, and nudges every Gaussian. Radiance-field training, splat speed.

  3. 03

    Keep the primitives

    A finished scene carries hundreds of thousands to millions of Gaussians — small enough to store, structured enough to query.

Ingestion

A bad frame rejected early is worth many rejected late.

Video decodes in hardware straight into GPU memory — no host round-trip. A frame gate scores blur and exposure on the decoder's own surface and rejects failures before a single byte is copied, because everything downstream is superlinear in frame count. What survives goes through pose recovery and training, then commits to the store. Captures queue behind one GPU worker.

  • Hardware decode to GPU
  • Blur & exposure gate
  • No host round-trip
  • One worker, queued captures
The ingestion path Video is decoded in hardware directly into GPU memory. A frame gate scores blur and exposure on the decoder's surface and rejects bad frames before they are copied anywhere; surviving frames go through pose recovery and training into the store. The host-memory lane above the pipeline stays empty — there is no round-trip. host memory — never visited gpu memory decode hardware frame gate blur · exposure pose + train one gpu worker store committed rejected before the copy capture capture capture queued behind the worker

Decomposition

One capture.
Taken apart, cleanly.

A scene is not one thing. Lumen separates a capture into branches — the static structure, the dynamic content that moved through it, and generated fill where nothing was seen — and can label primitives semantically by class. The espresso machine, the people and the room itself become separately addressable.

  • Static structure
  • Dynamic content
  • Generated fill
  • Semantic labels per class
A capture decomposed One captured scene separates into three stacked layers: the static structure of the room, the dynamic content that moved through it, and generated fill where nothing was observed. Primitives also carry semantic class labels, so parts of the scene — a machine, the people, the room — can be addressed separately. dynamic · what moved fill · generated static · the room espresso machine people the room

Storage

Splats stored like a database,
not like a file.

Every primitive is quantised to 27 bytes and appended to a write-ahead log, spatially ordered along a Hilbert curve, then compacted into immutable segments — write amplification fell from 44× to 1.1×, and opening a scene from 3.7 s of log replay to 103 ms. A level-of-detail pyramid and published quality tiers serve every reader, and read-at-sequence time travel makes any scene readable at any point in its history.

  • 27-byte primitives
  • Write-ahead log
  • Hilbert-ordered segments
  • LOD pyramid & quality tiers
  • Read at any sequence
The splat store Quantised 27-byte primitives append to a write-ahead log, are ordered along a Hilbert curve and compacted into immutable segments; a level-of-detail pyramid serves published quality tiers, and a sequence slider reads the scene at any point in its history. write-ahead log · 27 bytes per primitive compact hilbert order immutable segments lod pyramid published quality tiers read @ seq — any point in history
write amp 44× → 1.1× open 3.7 s → 103 ms

Key features

Places you can query.

Once a place is a scene in the store, it stops being footage and starts being an answer. Everything below runs against scenes Lumen has already built — and refuses honestly when it does not know.

Searching stored places Two text queries run against a store of Gaussian-splat scenes. A description of a kitchen counter finds the kitchen scene; a place the store does not know is refused rather than answered with a nearest guess. “a kitchen counter with pots and pans” “a pier at night” the scene store kitchen · scene 41 not in this archive

Find a place from a photo

Hand Lumen one picture and it returns the stored scene it was taken in — and where in it the camera stood.

Text to location

"A kitchen counter with pots and pans" finds the kitchen. A place the store has never seen gets an honest refusal, not a nearest guess.

Re-find automatically

When a new capture arrives, Lumen recognises the places it already knows and files the footage against them — no one tags anything.

Tell twins apart

Two near-identical hotel corridors, two branches of the same chain: geometry disambiguates places that photographs cannot.

Location scouting

Search an archive of places in words. The scout reads the answer instead of scrubbing the rushes.

Archive clustering

Group a lifetime of footage by place — measured at zero false merges, so a cluster never has to be untangled by hand.

Crowd-sourced venues

Verified captures from many people fuse into one scene of the venue. Geometric verification gates every contribution: zero false accepts.

Describe a scene in words

Splat to text: a vision-language model looks at a stored scene and writes down what the place is and what is in it.

25 min
to train a full scene on one GPU
103 ms
to open a scene that used to replay logs for 3.7 s
27 B
per stored primitive, quantised in the write-ahead log
0
false merges clustering an archive by place

Commercial offering

Lumen is commercial software.

Like Studio and Cloud, Lumen is one of the commercial ways to get MetaLoom — it is not open source and there is nothing to download here. It ships as a container image that runs on your own GPU hardware, next to the rest of your MetaLoom installation or on its own. If you have places on film and questions about them, write to us.

Delivery
container image, your GPU hardware
Licensing
commercial — like Studio & Cloud
Pricing
not published — ask us
Email us about Lumen Read the docs Join the Discord

An email reaches a person, not a funnel. There is no mailing list to add you to, so nothing happens to your address except a reply.