4D Gaussian Splatting (4DGS): Volumetric Video, Explained
Updated August 1, 2026 · The technology, the tools actually shipping, and what it costs to make a scene you can walk through in time
4D Gaussian Splatting (4DGS) adds time to 3D Gaussian Splatting: instead of reconstructing a frozen moment, it reconstructs a scene in motion — a photorealistic volumetric video you can pause, orbit, and walk through from any angle. Where a 3D splat is a still, a 4D splat is footage with a free camera.
Gracia shipped the first streamable 4DGS apps on Quest and Apple Vision Pro, Arcturus put splat sports replays in front of millions, and at SIGGRAPH 2026 4DV.ai, Dell, and Radiance Fields premiered an 11.5-minute film built entirely from 4DGS, the longest complete 4DGS production to date. This page explains how the technology works, what it takes to capture, and every tool you can actually use today.
What is 4D gaussian splatting?
A 3D Gaussian Splat represents a radiance field as millions of tiny translucent, colored blobs, Gaussians, that together reproduce photorealistic views from any position. 4DGS extends each of those primitives into the time dimension. Gaussians can move, deform, appear, and disappear, so the reconstruction is not just where everything is, but how it all changes across the length of a clip. Render any timestamp from any viewpoint and you get free viewpoint lifelike 3D video. This is the thing the volumetric video industry has chased for a decade.
A naming warning, because it trips everyone: two different seminal papers are both called “4DGS.” Wu et al.’s CVPR 2024 paper “4D Gaussian Splatting for Real-Time Dynamic Scene Rendering” deforms a static splat over time, while Yang et al.’s ICLR 2024 paper “Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting” uses genuinely four-dimensional primitives. When someone says “4DGS,” check which lineage they mean.
3DGS vs 4DGS at a glance
3D Gaussian Splatting
4D Gaussian Splatting
Captures
A frozen moment — one static scene
A scene over time, motion included
Output
One splat file (.ply, .spz, .sog)
A playable splat sequence or spacetime model
Capture hardware
One camera moved around the scene, or a lidar scanner
A synchronized multi-camera rig (20–60+ cameras) for production quality
Typical size
15 MB–2 GB per scene
Roughly 2–10 MB per second of footage when streamed
Where it plays
Browsers, game engines, headsets. Widely supported
Vendor apps, WebXR/WebGPU browsers, Unreal and Unity plugins
Standard formats
SPZ, SOG, glTF extension, USD support
None yet. Proprietary containers. MPEG has opened a splat-coding exploration
How 4DGS works: the four approaches
1. A static splat plus a deformation field
Train one canonical set of Gaussians, then a small neural network that bends them to each timestamp. One statue plus instructions for how it moves. This is Wu et al.’s 4D-GS, the most starred open source entry point. Compact and fast to train, but it struggles when objects appear or disappear mid-clip, and the whole model must download before playback, so it doesn’t stream.
2. Native four-dimensional Gaussians
Give every primitive a position and extent in spacetime, so Gaussians are born and die at specific moments and carry their own motion. Yang et al.’s ICLR 2024 formulation and OPPO’s Spacetime Gaussians established the idea. FreeTimeGS (CVPR 2025) pushed it to 450 fps at 1080p on a single RTX 4090, and its co-author Jiaming Sun is the CEO of 4DV.ai, which turned this line of research into a product.
3. Per frame splat sequences.
Reconstruct a splat per frame and play them back like video. Raw data is enormous, but it seeks, streams, and edits like footage. Gracia streams codec style splat sequences (keyframes plus motion deltas). 4DV.ai instead tracks Gaussians continuously through time, which suppresses popping artifacts, tolerates unsynchronized cameras, and enables what it calls infinite slow motion.
4. Feed forward and generative 4D
The research frontier of transformers and diffusion models that output dynamic Gaussians directly, with no per-scene optimization. NVIDIA’s BTimer reconstructs a bullet-time moment from casual phone video in about 150 milliseconds, and Meta’s 4DGT cuts reconstruction from hours to seconds. Fidelity still trails per scene training, but this is the fastest-moving corner of the field.
What it takes to capture
Production 4DGS still means a synchronized multi-camera rig. Research datasets use around 20 cameras. At NAB 2026, 4DV.ai and OBSBOT showed a 70-camera rig of 4K PTZ cameras, roughly $80,000 of commodity hardware, scalable to 200+ that compresses about 8 gigabits per second of raw feeds down to a 30–60 Mbps finished asset. 4DV’s InfiniteStudio capture system won Best in Show at SIGGRAPH 2025’s Real-Time Live, and portable rigs built from a handful of iPhones or GoPros now produce respectable casual captures.
Single camera 4DGS exists. Methods like MoSca lean on depth and tracking foundation models to fake the missing views. Additionally, people have extended Apple's SHARP to video, so they can create monocular 4D.
File sizes, streaming, and the missing standard
Raw per frame splat sequences run to hundreds of megabytes per second
Gracia streams at ~80 Mbps for full quality (120,000 splats per frame) with a 17 Mbps economy mode
There is no standard 4D splat format yet. Every vendor ships its own container, while static formats consolidate (SPZ, SOG, glTF’s new splat extension, OpenUSD support) and MPEG has opened a Gaussian splat exploration
The tools you can use today
Capture plus cloud service built on the FreeTimeGS lineage with continuous splat tracking, unsynchronized camera tolerance, WebXR playback on Vision Pro, Quest, and flat browsers. No self serve upload or livestreaming yet. Its showcase is the 11.5-minute all-4DGS film made with Dell and Radiance Fields, premiered at SIGGRAPH 2026. YouTube literally cannot play it, so it ships as WebXR and a Quest app. Company profile: 4DV.ai.
The distribution story of 2026. Gracia shipped the first true 4DGS streaming in March (keyframes plus motion deltas, 10× bandwidth reduction), then the first streamable 4DGS app on Apple Vision Pro in June, free, with 25+ scenes, alongside Quest, Pico, macOS, and Steam apps, Unity and Unreal plugins, and WebGPU browser playback. It also powers a photoreal 4DGS character at PortAventura World — a live theme-park deployment. Profile: Gracia.
Unreal Engine — Volinga and SplatRenderer
Volinga brought 4DGS into Unreal as a production workflow, with relighting, ACES/HDR, and virtual-production heritage via Disguise. On the free side, the open-source SplatRenderer plugin plays animated 4DGS sequences in UE 5.5+ at 100+ fps with full Level Sequencer keyframing.
Unity, web, and Blender
4Dviews unveiled a Unity plugin with native gaussian-splat export at AWE USA 2026, and UnityGaussianCapture added 4DGS support for training splats from moving Unity scenes. In the browser, ChronoSplat plays 4D splat sequences from static files — no server needed — and SuperSplat handles animated splats too. In Blender, KIRI Engine’s 3DGS Render 5.0 adds rigging, mocap retargeting onto splats, and 4DGS sequence export.
The volumetric video industry is switching to splats
The mesh-based volumetric industry spent a decade fighting hair, transparency, and uncanny skin. Splats solve exactly those failure modes, and the studios have noticed. 4Dviews’ HOLOSYS+ became the first capture system to export both mesh and gaussian splats natively; Evercoast now leads with splat reconstruction — it shot A$AP Rocky’s “Helicopter” video on a 56-camera rig — and Arcturus, steward of Microsoft’s Mixed Reality Capture technology, raised a seed round to chase live 3D sports broadcast, with its splat-based replay demos going viral in early 2026. Holdouts remain — Depthkit and most MRCS license studios are still mesh-first.
The milestone reel from the last twelve months: the first real-time-streamed 4DGS music performance (DNE × Gracia’s four-minute Amy May piece, 1.5 million opening-weekend views), a splat-native music video for A$AP Rocky, a live theme-park character, viral sports replays, and an 11.5-minute film. Costs are falling with it. Traditional volumetric ran five figures per finished minute on bespoke stages. 4DGS rigs are built from commodity cameras, though vendors still quote per-project.
FAQ
What is 4D gaussian splatting in one sentence?
It’s 3D Gaussian Splatting extended through time. A lifelike reconstruction of a radiance field in motion that you can watch from any viewpoint, like video with a free camera.
Is 4DGS the same thing as volumetric video?
4DGS is the technology. Volumetric video is the product category. Earlier volumetric video used textured meshes, which struggled with hair, transparency, and skin. 4DGS is now the highest quality way to make volumetric video, and most major studios are adopting it.
How big are 4DGS files?
Streamed, roughly 2–10 MB per second of footage: Gracia streams at about 80 Mbps full quality (with a 17 Mbps mode). Raw uncompressed sequences are hundreds of MB per second, and research codecs have squeezed benchmark scenes to ~11 KB per frame.
Can I capture 4DGS with my phone?
Not really — one moving camera can’t see a moving scene from all sides at once. Monocular research methods exist but are brittle. Practical minimum is a small synchronized rig (people have used ~10 iPhones or GoPros); production work uses 20–60+ cameras. However, it should be likely that a 4D equivalent of World Labs will emerge over the next year.
What can play 4DGS today?
Quest 3 and Apple Vision Pro (Gracia’s free apps, plus WebXR), desktop and mobile browsers via WebGPU players like ChronoSplat and SuperSplat, and Unreal or Unity through Volinga, SplatRenderer, and the 4Dviews and Gracia plugins. There’s no YouTube for 4DGS yet.
How much does 4DGS volumetric video cost to produce?
Vendors quote per project, so take ballparks: traditional mesh volumetric ran five figures per finished minute on six-figure stages, while 4DGS capture rigs are commodity hardware (the 60-camera NAB 2026 rig was roughly $60,000 of cameras) plus cloud GPU processing. Costs are falling fast, but nobody publishes a rate card yet.