Believe the hype or don’t, but read the invoice first. When A$AP Rocky dropped the Helicopter video, the internet argued about whether it was AI. It wasn’t — as the team at radiancefields.com reported, nearly every human figure, Rocky included, was captured volumetrically on Evercoast’s rig and rendered as dynamic Gaussian splats. Over 10 terabytes of raw data went in; roughly 30 minutes of splatted footage came out as about one terabyte of PLY sequences. I am the money that clears those render bills, and I want to teach you what a splat actually is — because the concept underneath that video is about to reprice a line on every Swiss studio’s survey invoice.

←TODAY: In 2026 a five-minute phone walk-around already yields a metric-accurate 3D capture; a music video proves the same maths captures a moving body. →3012: As-built reality becomes a fitted field you query, not a mesh you rebuild — survey becomes a subscription. Fulcrum: The technique got cheap the moment rendering moved from ray-march to rasterisation — so the value drained off the render farm and pooled on the consumer GPU under your desk.

The concept, not the celebrity

What it is: A 3D Gaussian Splatting model is a scene stored as a cloud of about a million little translucent blobs. Each Gaussian carries a mean position μ in space, a 3×3 covariance Σ that stretches and rotates it into an ellipsoid, an opacity α, and a view-dependent colour held as spherical-harmonic coefficients up to degree three. To render a frame you project each blob to the image plane, depth-sort per tile, and alpha-composite — a rasterisation, not a physics march through a volume. That single choice is why the pipeline saturates a consumer GPU at 100-plus frames per second at 1080p. The scene is not a model of the building; the blobs, fitted to your photos, are the building.

Why it works: The lineage runs through the emission-absorption equation Kajiya and Von Herzen formalised for clouds and smoke in 1984 — how much light a semi-transparent medium emits and absorbs along a ray. NeRF wrapped that integral in a small multilayer perceptron and let gradient descent fit the network to a hundred photos at once; PAZ’s own concept panel on radiance fields puts it exactly: the weights of that network are, literally, the building. Splatting keeps the physics but throws out the implicit network for explicit, differentiable blobs you can rasterise. Two languages for one captured reality — an implicit field is a continuous light-transport model, a Gaussian splat is its rasterisable twin. The economic consequence is blunt: NeRF’s frames took seconds and its training took hours of rentable compute; splatting runs in real time on a card you already own. The capex drained off the render farm and pooled in a one-time purchase. Every technique that makes the frontier run on a consumer GPU quietly redraws who gets paid.

Origins: Two acts, both datable. In March 2020, Ben Mildenhall, Pratul Srinivasan, Matthew Tancik, Jonathan Barron, Ravi Ramamoorthi and Ren Ng presented NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis at ECCV — a five-dimensional coordinate in, a colour and a density out. Act two landed at SIGGRAPH 2023, when Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler and George Drettakis at INRIA Sophia-Antipolis published 3D Gaussian Splatting for Real-Time Radiance Field Rendering, replacing the network with millions of splats and a custom CUDA rasteriser at more than 100 frames per second at 1080p. Mittal’s 2024 survey already catalogues more than five hundred NeRF preprints, the field roughly doubling each year. A$AP Rocky, incidentally, is not new to this — his 2023 Shittin’ Me video showed several NeRFs and even the Instant-NGP GUI on screen. The novelty in Helicopter is dynamic splatting: performance, not just architecture, preserved spatially.

Read the money, not the round

Here is where I earn my keep. Per radiancefields.com, the Helicopter shoot in Los Angeles this August deployed a 56-camera RGB-D array synchronised across two Dell workstations, recorded more than 10 TB of raw data, and exported around 1 TB of PLY sequences that then flowed through Houdini with CG Nomads’ GSOPs and OTOY’s OctaneRender for relighting. Chris Rutledge of Grin Machine, the project’s CG Supervisor, framed the reason plainly — the same freedom would have been impractical or prohibitively expensive on a conventional VFX pipeline. That is the whole trade: you spend once on capture and storage to buy unlimited, cheap creative decisions downstream. It is a workflow that resembles simulation more than filming, and simulation has a very different cost curve — heavy fixed capex, near-zero marginal iterations.

For an architect the same curve reappears, smaller and closer to your invoice. A splat capture of one building corner, per PAZ’s radiance-field panel, is good enough for as-built verification, heritage documentation, façade survey and clash-checking against a real building rather than a CAD wish. The value that used to pool with a surveyor — half a day, tape and total station — now pools with whoever owns the rig and the GPU-hours. But watch the drain: 1 TB per 30 minutes of footage is not a metaphor, it is a storage bill that runs forever. The capture is cheap; the keeping is the meter that never stops.

In practice: A Zurich studio does not need a 56-camera array to touch this. The Monday move is smaller and denominated in francs. Take the phone you already own, capture one contested corner — a listed façade before a Wettbewerb submission, an existing condition your Bauleitung disputes — fit a splat with the INRIA reference implementation overnight on the office workstation, and measure the geometry off the fitted field instead of scheduling a survey. Set one policy alongside it: a firm rule on where the PLY sequences live and how long you pay to keep them, because uncatalogued captures become a storage subscription nobody signed off. The office that treats a splat as an asset with a carrying cost — not a free souvenir — is the one whose balance sheet still closes next year.

Hack: Find the scan count where the capture rig stops costing you money and starts making it — before the GPU generation you bought it on is written off. The frontier is exciting; the break-even is what decides whether you get to keep doing it. Three or four numbers, one division, and you know your runway. Change the constants to your own quote and re-run.

capex, price, op_cost, life_m = 180_000, 800, 120, 24   # CHF; life in months
margin    = price - op_cost              # each scan's contribution
breakeven = capex / margin               # scans to earn the rig back
pace      = breakeven / life_m           # scans/month before write-off
print(f"break-even {breakeven:.0f} scans -> {pace:.1f}/month for {life_m} mo")

Run it. If the pace it prints is higher than the work your pipeline can plausibly bill, the rig is a subsidised hobby, not a service line — and better to know that in Python this week than in your annual accounts next spring.

What a survivor packs

One warning from the far side. The crashes I remember clearing were never the tools that admitted they cost money — those got priced and drained cleanly. The damage came from the venture-priced proptech stacks that called nine subsidised quarters “traction,” got poured into critical office workflows, and went dark together the quarter the subsidy expired. Splatting itself is durable — it is INRIA maths on a consumer GPU, it does not need anyone’s runway. The platforms wrapping it in a monthly “capture cloud” are the ones to read carefully. Before you commit your survey practice to one this quarter, find the number that says how it earns back its own capex. If no one will show you that number, that is the number.

So: capture the corner Monday, fit the splat, measure it against a tape, and run the break-even on your own quote before you sign anything with a monthly fee attached.

Source: Hacker News