4D Gaussian Splatting Currently, it is one of the most exciting topics surrounding volumetric media. The idea is immediately fascinating: not only capturing a room or an object as a photorealistic Gaussian Splat, but also motion – people, performances, sports, clothing, or entire dynamic scenes.
And this is where an interesting misunderstanding is currently emerging: Anyone who watches current demos quickly gets the impression that the technology should already be ready for production. After all, some examples look spectacular, can be viewed freely, and run in real time.
But between a convincing 20- or 30-second demo and a media format that can be streamed reliably for several minutes, there is still a considerable technical gap. The bottleneck is now less the image quality than the latency. Data volume, compression, streaming, capture effort, and standardization.
In this article, we look at what 4DGS actually is, which systems are already functioning today, and why the technology is already impressive in its scope and still surprisingly immature.
What is 4D Gaussian Splatting?
In the classic 3D Gaussian Splatting A static scene is described by very many small three-dimensional Gaussians. These contain, among other things, information about position, size, orientation, color, and transparency.
With 4DGS comes the Time as an additional dimension.
A person raises their arm. Clothing deforms. Hair moves. Body parts cover one another and become visible again later. A 4DGS system therefore not only has to describe where something is, but also how this spatial representation changes over time.
What is important here: 4D Gaussian Splatting is not currently a uniform file format. The term encompasses various research approaches. Some work with a static initial scene and a time-dependent deformation model, others with Gaussian trajectories, keyframes, or directly with four-dimensional primitives.
4DGS is therefore a rather Technology family as an established media standard.
Why is 4DGS so interesting?
4DGS could in the long term bridge the gap between traditional video and real 3D.
In a normal video, the perspective of the original camera remains intact. In a volumetrically captured 4D scene, however, the viewer can move freely within certain limits within the scene.
This opens up interesting applications in:
- Film and VFX
- Sport and Performance Capture
- Dance and Theater
- virtual concerts
- XR and Spatial Computing
- Training and Simulation
- Free-viewpoint video
- Games and Virtual Production
The presentation is particularly exciting for XR: A real human being is not presented to us as a flat video, but actually physically present in the scene.
The persistent myth: „It really should already work“
Around 4D Gaussian Splatting there is currently a noticeable perception gap. Many current demos look so convincing that the question quickly arises:
„If that’s the case – why can’t we just produce several minutes of it and stream it like a video?“
This is where one of the most important misunderstandings lies.
Real-time rendering is not the same as real-time streaming
The fact that an already loaded 4DGS dataset with high frame rate can be displayed on a powerful GPU does not say anything about the size of this dataset, how long the download takes, or how much memory and bandwidth it requires.
With video, we have become accustomed to multiple minutes of high-resolution material starting almost immediately. However, behind this lies decades of development in codecs, hardware decoders, keyframes, adaptive bitrates, and streaming infrastructure.
This infrastructure is just being created at 4DGS.
An up-to-date research paper on PD-4DGS Many existing 4DGS representations are still described as monolithic bitstreams that must be transmitted completely or largely first. The progressive method presented there reduces the first-frame latency in a 2-Mbit/s test from previously 73 to 930 seconds to around 1.7 seconds.
The very fact that such fundamental streaming problems are still the subject of current research in 2026 shows very well where the technology currently stands.
Why are so many 4DGS demos so short?
Not every short demo is therefore short because a longer sequence would not be technically possible. Nevertheless, it is striking that many publicly displayed 4DGS examples are compact sequences of a few seconds to a few tens of seconds.
This is understandable, because with each additional second, depending on the procedure, among other things:
- Capture data volume
- Reconstruction costs
- Processing time
- Storage requirement
- Download size
- Decoding effort
- GPU and RAM requirements
An impressive 20-second demo is therefore technologically something completely different from a five-minute volumetric film that starts instantly on a normal device via a typical internet connection.
The demo is therefore not automatically production yet.
The big challenge: data volume and temporal redundancy
With a static Gaussian Splat, a state of a scene must be stored. With a dynamic scene, the changes over time must also be described.
The simplest, but most inefficient approach would be to store a complete new Gaussian Splat for each frame. Modern methods therefore try to exploit redundancies: for example, they store a baseline and then only describe how Gaussians move or change.
This conceptually reminds of modern video compression. The difference is, however: For video, there are established standards such as H.264, H.265, or AV1. For 4DGS, there is no comparable, generally accepted codec yet.
Compression already exists – but not yet a common standard
It would be incorrect to say that there is no compression for 4DGS. On the contrary: Compression is one of the most active research fields in this area.
Current approaches reduce data by, among other things:
- Pruning redundant Gaussians
- Quantization of parameters
- Time forecast
- Keyframes and Deltas
- Simplified trajectories
- Entropy coding
For example, describe P-4DGS a predictive approach that utilizes spatial and temporal redundancy similar to a video codec and achieves very high compression rates in the published tests.
Such research results are promising, but they cannot easily be compared with one another. Data sets, quality targets, and baseline models differ significantly.
The actual question now is no longer: Can 4DGS be compressed? Rather: Which approach will be interoperable, sustainable, and a standard in the long term?
Gracia AI: A commercial end-to-end pipeline
Gracia AI Currently one of the most visible companies to offer 4D Gaussian Splatting as a complete production pipeline.
The workflow includes capture through partner studios, processing, cleanup, compression, and distribution. Gracia currently speaks of a 45-fold near-lossless compression and offers expenses for streaming/web, VFX, and XR headsets.
It is particularly interesting that Gracia 4DGS is now streaming directly via WebXR and providing SDKs for different target platforms.
However, this is a proprietary system. The .mint format itself and the internal compression pipeline are not an open industry standard that any other tool could simply implement.
Gracia thus demonstrates very convincingly that the problem is fundamentally solvable – but not yet via a generally established pipeline.
4DV.ai: 4D Video as a complete media system
The Chinese company as well 4DV.ai is working on a complete creation and distribution pipeline for 4D video based on Gaussian Splatting.
4DV.ai explicitly positions 4D video as a new visual medium for film, music, sports, fashion, and advertising. The company received the Best-in-Show award at SIGGRAPH Real-Time Live! in 2025 for a complete 4D video workflow.
It is also noteworthy that 4DV.ai, together with research partners, has developed its own Volumetric Video Compression Challenge It was evaluated there, among other things, model size, decode time, and rendering performance.
This shows very clearly: Even for companies that already have far-advanced 4DGS demos, efficient compression remains one of the key development areas.
PlayCanvas: 4DGS comes in a real web engine
What is particularly interesting is the integration of Gracia-4DGS into the open-source web engine. PlayCanvas.
This allows for the dynamic playback of 4DGS content, as well as its use within a real 3D scene. The integration supports, among other things:
- Real-Time Relighting
- Shadow Casting
- Post-processing
- mixed scenes from 4DGS, static 3DGS splats, and classic meshes
This is particularly exciting because PlayCanvas already has an unusually deep native integration in static Gaussian Splatting. For 3DGS, webGPU rendering, streaming, LOD, and compressed formats are now available.
4DGS is not yet a universal standard there – but the integration shows very concretely how dynamic splats could become part of normal real-time and web pipelines in the future.
Try it out for yourself: Current 4DGS demos and projects
Anyone who wants to get a feel for the current state of the technology should definitely try it out in person. This way, you can assess charging time, quality, and interactivity much more clearly than through demo videos.
- PlayCanvas + Gracia Live Demo: demo.gracia.ai/playcanvas.html
4DGS in the browser with Relighting, shadows, and integration into a PlayCanvas scene. - Gracia AI: gracia.ai
Commercial 4DGS pipeline for web, XR, and VFX with several interactive examples. - 4DV.ai: 4dv.ai
Examples of 4D video, capture and distribution from China. - Open .4dgs format: GitHub: avala-ai/4dgs
An interesting attempt to define a streamable and seekable open container format for 4DGS. - Unity 4DGS Playground: ITHACA 4DGS Unity
Experimental Unity integration for temporal Gaussian splats.
These examples in particular show the current state very well: The visual results are impressive, but depending on the demo, loading times, hardware requirements, and proprietary dependencies become quickly apparent.
The capture process remains challenging as well.
In static Gaussian Splatting, ideally nothing moves. In 4DGS, however, motion is the real content.
This creates additional challenges:
- Synchronization of multiple cameras
- Coverings between body parts
- Quick movements
- Hair and clothing
- Hands and fingers
- changing reflections
- Lighting consistency
High-quality systems therefore often rely on fully synchronized multi-camera rigs. Gracia mentions setups with around 60 cameras and hardware synchronization or genlock for full-360 capture, for example.
Portable rigs reduce the effort and make on-location capture possible, but they naturally provide less controlled conditions than a permanently installed studio.
The most likely short-term deployment: Hybrid worlds
For many productions, it is probably not necessary to build a complete scene as 4DGS.
A very sensible approach is:
Static 3DGS Environment + 4DGS Character
The space is captured once as a high-quality static Gaussian Splat. Only the actually moving person is recorded as 4DGS.
Alternatively, Gaussian splat environments can be combined with classic 3D characters, video elements, or individual volumetric performances.
Especially in XR and virtual production, this hybridization is likely to be significantly more practical in the short term than fully dynamic 4DGS worlds.
Where are we in 2026?
| Area | As of 2026 |
|---|---|
| Visual Quality | Already very convincing |
| Real-time rendering | Basically possible |
| Capture | Works, but often hardware-intensive |
| Compression | Active research and first commercial solutions |
| Streaming | First functioning systems, but not yet a general standard |
| Web | Gracia/PlayCanvas showcase first convincing integrations |
| Unity | Possible with experimental plugins |
| 4DGS file format | First open approaches, not yet established |
| Interoperability | One of the largest open construction sites |
Conclusion: The images are already there – what’s missing now is the infrastructure
4D Gaussian Splatting is currently one of the most exciting developments in the field of volumetric media.
The visual side of the technology is already astonishingly far along. Research systems and commercial providers are showing moving people, performance capture, and freely viewable dynamic scenes.
But it is precisely these current demos that also show very clearly the limits: The transition from a short, spectacular sequence to a medium that can be streamed reliably over multiple platforms and across multiple platforms is still very significant.
Companies like Gracia AI and 4DV.ai are already solving parts of this pipeline. PlayCanvas shows how 4DGS can be integrated into a modern web engine. At the same time, initial open format approaches and research projects for progressive transmission and compression are emerging.
What is still lacking is a common ecosystem of Codec, container format, streaming, LOD, editing, and runtime support.
The breakthrough of 4DGS will therefore probably not come through even better images – but through the codec and the infrastructure that really make these images practical.
Those who would like to start with classical Gaussian splatting can also find it here. Gaussian-Splatting Beginner's Guide, our overview Tools for Gaussian Splatting as well as our article about Gaussian Splatting in the Browser.
FAQ about 4D Gaussian Splatting
4D Gaussian Splatting extends static 3D Gaussian Splatting with a time dimension. This allows for the volumetric capture and viewing of moving people, objects, or entire dynamic scenes from different perspectives.
Yes for specialized productions, but not yet as a generally standardized media format. Companies like Gracia AI are already offering commercial pipelines. However, compression, streaming, file formats, and cross-platform playback remain highly fragmented.
In addition to spatial information, 4DGS must also store changes over time. Motion, deformation, visibility, and appearance increase the data volume significantly compared to a static Gaussian Splat.
Yes. Both research and commercial providers use different compression methods. However, there is still no universally accepted codec that would be comparable to H.264, H.265, or AV1 for conventional video.
Yes. Gracia AI offers web streaming, and an integration with PlayCanvas demonstrates 4DGS directly in a web engine, including relighting, shadows, and post-processing.
Not natively a standard feature of the engine. However, there are experimental and community-based plugins that allow dynamic Gaussian splats to be displayed and played in Unity.
High-quality productions often use synchronized multi-camera systems. Depending on the quality requirements, the setups range from portable camera rigs to complete 360° studios with dozens of cameras.
Not just the pure rendering quality anymore, but the entire media pipeline: efficient compression, progressive transmission, streaming, editing, standardized file formats, and reliable playback on different end devices.
