How 4dv.ai Works: The Technology, the Team, and Why It's All Demos

Disclaimer: This article aggregates publicly reported research, papers, and vendor/press announcements. Performance figures (training speed, data rates, FPS, compression, camera scaling) are vendor or paper claims, not independent benchmarks; some company/funding details could not be confirmed in public registries. The reasoning involved AI-assisted generation, has not undergone peer review, and may contain errors. Data are current as of June 2026.

TL;DR

4dv.ai (Chinese brand "视维智能," 4DV Intelligence) is a volumetric-video company that captures real moving scenes with multi-camera rigs and reconstructs them as "4D Gaussian Splatting" assets — millions of colored, semi-transparent 3D blobs (Gaussians) that move over time, letting viewers freely move the camera (6 degrees of freedom) inside a recorded video, in a web browser.
The core technology is the academically published method FreeTimeGS (CVPR 2025) and its lineage (4K4D, EasyVolcap) from Zhejiang University's 3D Vision Group, co-authored by 4DV founder/CEO Jiaming Sun (孙佳明). Because the method is fully public in peer-reviewed papers, the company website has no need to re-explain it.
The website shows only demos because 4DV is a commercially-focused, early-stage startup: the science lives in arXiv papers, while the website targets clients, partners, and investors. The shipping pipeline (capture → cloud processing → streamed 4D asset) and the production engineering around it are the actual trade secrets — and demos are the most persuasive sales artifact for a visual medium.

Key Findings

What 4dv.ai is. 4DV.ai's own site states: "we are working on 4D video creation, playing and editing tools based on Gaussian Splatting. We envision 4D video to be the future of visual media and ultimate medium for immersive storytelling." It is a volumetric / 6DoF / free-viewpoint video company. "4D" = 3D space + time. The output is a dynamic scene you can walk around inside, viewable on flat screens, Apple Vision Pro, Meta Quest, and PICO via WebXR.

Who is behind it. Founder and CEO Jiaming Sun (孙佳明) is affiliated with the 3D Vision Group (ZJU3DV) at Zhejiang University, led by Prof. Xiaowei Zhou. Sun previously worked at DJI and SenseTime and co-founded IDR (Image Derivative Inc. / 杭州像衍科技有限公司, registered 2021-09-24, legal rep Guo Yudong / 郭玉东). The Chinese brand is 视维智能 ("4DV Intelligence"), based in Hangzhou with a US/LA footprint. IDR — the predecessor company, whose shareholders include ZJU professors Bao Hujun (鲍虎军) and Xu Weiwei (许威威) — raised an undisclosed angel round (announced ~January 2022) from IDG Capital, Sequoia China Seed Fund (红杉中国种子基金), and Trilobite Capital (三叶虫创投). No separate funding round has been publicly announced for the 4DV entity itself, and no registry record for the 4DV legal entity could be confirmed.

The underlying technical method. The core algorithm is FreeTimeGS (CVPR 2025; arXiv:2506.05348), authored by Yifan Wang, Peishan Yang, Zhen Xu, Jiaming Sun, Zhanhua Zhang, Yong Chen, Hujun Bao, Sida Peng, and Xiaowei Zhou (Zhejiang University and Geely Automobile Research Institute). It builds on the group's earlier 4K4D (CVPR 2024) and the EasyVolcap framework (SIGGRAPH Asia 2023).

Recognition. 4DV's productized system "InfiniteStudio: 4D Volumetric Capture for Film Making and Beyond" won Best in Show at SIGGRAPH 2025 Real-Time Live!. Per SIGGRAPH's official PR (PR Newswire, August 14, 2025), it was credited to "Jiaming Sun, Siyu Zhang, and Yu Zhang with 4DV.ai; Zhi Wang with MediaStorm; Yuke Ding; and Zhen Xu and Xiaowei Zhou with Zhejiang University." 4DV states it is the first Chinese team to win in the program's decade-long history.

Details

What the product actually does

Traditional video locks you to the camera operator's single viewpoint. 4DV captures a real scene with many synchronized cameras and reconstructs it into a representation you can navigate freely in space and time. The viewer runs in a normal web browser using the PlayCanvas/WebGL engine, with WebXR support, so a headset browser (Vision Pro, Quest, PICO) can render it in stereo with 6DoF — you can lean and look around occluded areas, not merely rotate your head in place.

UploadVR tested a public WebXR demo built from about 20 camera views and described the result as "like a 180° 3D video but with 6DoF." On data efficiency, UploadVR reported: "4DV AI's data efficiency is also impressive, with each second of footage taking around 12.5 megabytes, meaning in theory for a 100 megabit per second internet connection it would take a second of loading time per second of footage." On training speed, UploadVR noted: "While Gracia shipped a moving splats demo last year, that system took six minutes to train a single frame. In contrast, 4DV's paper claims its technology is 30 times faster, training a whole second of video in that time." Both are vendor/paper claims.

The technical pipeline, step by step

Multi-camera capture. A synchronized array of cameras records the scene. Public demos used ~20 cameras covering a frontal arc. Per OBSBOT's April 20, 2026 press release, the NAB Show 2026 demonstration (joint Booth C5249, Las Vegas) used "a synchronized array of approximately 60 OBSBOT Tail 2 cameras [that] feeds 4DV.ai's real-time 4D capture and reconstruction pipeline. The architecture is designed to scale to 200 or more cameras in full production deployments." Per CineD, each Tail 2 produces roughly 120 Mbps, so a ~70-camera array generates around 8.4 Gbps of raw footage.
Calibration & preprocessing. Camera poses are estimated; the system requires precise per-frame calibration for every camera. Data is gathered to a central server and uploaded to 4DV's data center.
4D reconstruction (the core). A 4D Gaussian Splatting model is fitted. In FreeTimeGS, Gaussian primitives are allowed to appear at arbitrary times and locations — rather than being defined in a canonical space and warped by a deformation field, an approach that struggles with fast/complex motion. The paper's abstract explains: "we propose FreeTimeGS, a novel 4D representation that allows Gaussian primitives to appear at arbitrary time and locations… we endow each Gaussian primitive with an motion function, allowing it to move to neighboring regions over time, which reduces the temporal redundancy." A temporal opacity function controls when each Gaussian is "active." Training uses a 4D regularization loss, periodic relocation of primitives, and a rendering loss.
Compression. Because the continuous representation shares splat data across time rather than storing each frame independently, the data compresses dramatically. Reported figures vary by source: CineD's interview cites the final 4DGS output compressing to roughly 30–60 Mbps for the OBSBOT/NAB rig, while the InfiniteStudio Real-Time Live! paper (ACM, doi 10.1145/3721243.3735979) cites "high compression rates of 80–120 Mbps" (abstract) and "100–200 Mbps" (body). Either way, this is low enough to stream over a normal home internet connection.
Real-time rendering / delivery. The result is delivered as a 4D asset and rendered in real time. FreeTimeGS reports real-time rendering "at 1080p resolution with a speed of 450 FPS using a single RTX 4090 GPU." The browser viewer uses PlayCanvas/WebGL; the research stack uses the EasyVolcap framework and a fast Gaussian rasterizer (zju3dv/fast-gaussian-rasterization).

How this differs from "vanilla" 4D Gaussian Splatting

Per the CineD interview with Sun, traditional 4DGS pipelines reconstruct the scene frame by frame, with each frame's splats computed independently — which produces visible flickering, because no information is shared across time, and ignores the fact that most of the scene is highly redundant from one moment to the next. 4DV instead tracks the splats across time as a continuous representation. This both reduces flicker and enables the strong compression that makes streaming feasible. This temporal continuity is the central differentiator FreeTimeGS provides over methods like the original 4DGS and Spacetime Gaussians (STGS); the paper reports PSNR improvements of 2.4 dB over 4DGS and 1.4 dB over STGS on its SelfCap dataset (4.1 dB and 2.6 dB in dynamic regions).

The research lineage (why it's technically credible)

EasyVolcap (SIGGRAPH Asia 2023) — the group's open-source volumetric-video research framework (Xu, Xie, Peng, Lin, Shuai, Yu, He, Sun, Bao, Zhou).
4K4D (CVPR 2024) — a 4D point-cloud representation with hardware rasterization. Per the paper: "our representation can be rendered at over 400 FPS on the DNA-Rendering dataset at 1080p resolution and 80 FPS on the ENeRF-Outdoor dataset at 4K resolution using an RTX 4090 GPU, which is 30× faster than previous methods."
FreeTimeGS (CVPR 2025) — the current core method, outperforming prior 4DGS/STGS approaches on complex-motion scenes.
InfiniteStudio (SIGGRAPH 2025 Real-Time Live!) — the productized capture system, integrated with Unreal Engine via a plugin, that won Best in Show.

Jiaming Sun is a co-author across this lineage and of other well-known ZJU3DV works (LoFTR, NeuralRecon — a CVPR 2021 Best Paper candidate — and OnePose).

Commercial activity

4DV has shown completed productions spanning very different scales — from a small cat to a 13-person stage performance to a rocket launch — most experienceable as interactive demos on the site. It collaborated with MediaStorm (影视飓风) on iPhone 17-related content, executing virtual camera moves impossible with a physical rig, and on an ASUS 4DGS project at the Yungang Grottoes (云冈石窟). It partnered with OBSBOT at NAB Show 2026 to jointly develop a scalable multi-camera capture rig (OBSBOT supplying imaging hardware, 4DV supplying the software stack), and it organized the 1st Volumetric Video Challenge at SIGGRAPH Asia 2025 (Hong Kong, December 16, 2025). Sun told CineD the company is past research mode and ready to ship; live streaming is not yet part of the product — the current workflow is capture-then-process, delivering a 4D asset rather than a real-time feed.

Why the website only shows demos

This is the most analytically interesting question, and several reasons converge — with the most important one often overlooked:

The science is already public, so there is nothing to hide on the website. The underlying methods (FreeTimeGS, 4K4D, EasyVolcap) are published at top venues (CVPR, SIGGRAPH Asia) with arXiv papers, project pages, and open-source code. A potential client or investor doesn't need a technology explainer on 4dv.ai — the definitive explanation is the peer-reviewed paper. A research write-up on a marketing site would be redundant.
The website's audience is buyers, not researchers. The site funnels visitors to "商务合作" (business cooperation) and recruiting forms. Demos are the most persuasive possible artifact for a visual medium — "step inside the video" sells far better than equations.
The genuine trade secrets are in the production engineering, not the published core algorithm. The hard, defensible, un-published parts are the capture-rig design and synchronization, multi-camera calibration at scale, the cloud reconstruction pipeline, the compression/streaming stack, and the editing tools. These are precisely what a marketing site would never document — and what differentiates a shipping product from a published paper.
Commercial / competitive stage. 4DV is early-stage, transitioning from research to product. Demos build buzz (it went viral via UploadVR and Product Hunt) while the self-serve product and live pipeline remain in development; multiple third-party reviews note it currently operates as a demonstration platform with pre-processed content.

In short, the "demos but no docs" pattern is not evasiveness. It is the natural result of a research-heavy team that publishes its science in academia, keeps its production engineering proprietary, and uses its website purely for commercial demonstration.

Recommendations

To understand the technology deeply (do this first): Read FreeTimeGS (arXiv:2506.05348) and its project page (zju3dv.github.io/freetimegs), then 4K4D (arXiv:2310.11448) and EasyVolcap (arXiv:2312.06575). The EasyVolcap and 4K4D GitHub repos (github.com/zju3dv) contain runnable code reflecting this family of methods. These are the authoritative sources — vastly more informative than the website will ever be.

To evaluate it commercially: Try the WebXR demos directly (e.g., the salmon, corgi, and hair scenes linked from the FreeTimeGS page and 4dv.ai) on a Vision Pro or Quest, then use the "商务合作" Typeform to request capabilities, pricing, and turnaround. Benchmark against alternatives — Gracia, Varjo Teleport, Niantic's Scaniverse, and frame-independent 4DGS pipelines — paying attention to flicker, occlusion artifacts, and file size.

Thresholds that would change the assessment: If 4DV ships (a) a self-serve upload-to-4D pipeline, (b) a live 4D streaming product, or (c) an announced institutional funding round, it has moved from "impressive research demo" to "scaling product" and warrants re-evaluation as a serious commercial vendor. Conversely, if capture remains studio/rig-bound with multi-hour cloud processing, it stays a high-end production service rather than a mass-market tool — judge it accordingly.

Caveats

Performance figures (30× faster training, ~12.5 MB/s footage, 450 FPS at 1080p, compression to 30–60 Mbps or 80–200 Mbps depending on source, 200+ camera scaling) are vendor or paper claims, not independent benchmarks. The two compression figures (CineD's 30–60 Mbps vs. the InfiniteStudio paper's 80–200 Mbps) come from different contexts and should both be treated as published targets.
No registered legal entity record, registration date, or funding round for the 4DV/视维智能 company itself could be confirmed in public registries. The confirmed angel round (IDG Capital, Sequoia China Seed Fund, Trilobite) belongs to the predecessor company IDR/像衍科技 (杭州像衍科技有限公司), not to 4DV.
The precise division of labor on the MediaStorm iPhone 17 work — exactly which footage was processed through 4DV's 4DGS pipeline versus MediaStorm's own bullet-time rig — is not fully documented.
Many third-party "AI tool review" sites describe 4dv.ai as converting a single 2D video into 4D. This is inaccurate marketing paraphrase: the real pipeline is multi-camera capture, consistent with the academic papers and the founder's own description. Treat those review sites as unreliable on technical specifics.