Multi-camera video → 4DGS
Convert multi-camera video to 4D Gaussian Splatting online
4dgspl.at is a self-serve cloud service: upload synchronized videos of the same moment from 4–8 cameras plus their calibration, and the moving scene is rebuilt as 4D Gaussian Splatting (4DGS, also called dynamic Gaussian splatting). You get a scene that plays in real time in the browser and that viewers can orbit, zoom and pause from any angle. No GPU, no Python environment, no research code to compile.
What you get
- Browser player: the 4DGS sequence plays on the project page (PlayCanvas, WebGL); drag to rotate, scroll or pinch to zoom, scrub the timeline.
- Share link and embed code: a public link that needs no sign-in and an
<iframe>for your own website. Links can be closed at any time; your input videos are never shared. - Files: one ZIP with a 3D Gaussian Splatting file per frame (
compressed.ply),sahne.json(frame count, fps, camera, quality metrics) and an MP4 preview. A single frame opens in SuperSplat or any 3DGS viewer.
What you need to upload
| Item | Requirement |
|---|---|
| Cameras | At least 4, 8 recommended (6 minimum for clean results); on a ring around the subject at equal angles; with 8+ cameras use two heights |
| Videos | cam00.mp4, cam01.mp4 … in ring order; same resolution, frame rate and start frame; cameras never move during the take |
| Calibration | A ZIP with COLMAP cameras.txt (PINHOLE or SIMPLE_PINHOLE) and images.txt, image names cam00.jpg … matching the videos; or 4dgs-cameras JSON (OpenCV convention) |
| Subject and background | The subject fills 70–85 % of the frame height; plain backdrop (pure black in virtual scenes); constant, soft light |
| Length | The first 4 seconds of each capture are processed today, at 15 fps (60 frames) |
Where the footage comes from does not matter: a volumetric studio rig, 4–8 smartphones on tripods, or virtual cameras in Blender, Unreal Engine or Unity. The capture guide has the phone procedure (locked settings, clap sync, ffmpeg alignment) and the COLMAP commands; Blender render to 4DGS covers virtual rigs.
How it works
- CaptureRecord the same moment from every camera. Give the scene a few seconds of textured objects at the start if you will calibrate with COLMAP.
- Sync and nameTrim all videos to the same first frame and export them as
cam00.mp4… with the same resolution and frame rate. - CalibrateRun COLMAP on one frame per camera (or export the poses of your virtual cameras) and zip
cameras.txt+images.txt. Check the folder with the free delivery checker. - UploadCreate a project in the dashboard, drop each video into its camera slot, add the calibration, pick Fast, High or Highest quality and submit. Coding agents can do the same with the upload script and agent key.
- ProcessingA worker on our GPU server (NVIDIA RTX 5080) reconstructs every frame as a 3D Gaussian Splatting scene with the open-source gsplat library. The project page shows the queue position, the stage and live progress.
- Watch and sharePlay the result, open a share link, copy the embed code or download the ZIP.
Ways to get from multi-camera video to 4DGS
| Approach | Good for | What it takes |
|---|---|---|
| Open-source research code (4DGS papers, gsplat-based pipelines) | Researchers who want to change the method | A CUDA GPU, a Python/CUDA environment, data preparation scripts and time to debug them |
| Managed capture studios and vendors | Large productions, live events, sports | Booking their rig and team; quotes per project |
| 4dgspl.at | Creators, studios and developers who already have, or can render, multi-camera footage | A browser and synchronized videos with calibration; results in the browser, as links and as PLY files |
Pricing
Sign up with email or Google. Every new account starts with 5 free jobs of up to 10 seconds each at Fast quality. After that, credits are used: 1 credit = 1 second at Fast quality, High ×3, Highest ×8. Credit packs are coming soon.
Frequently asked questions
Can I make 4D Gaussian Splatting from a single video?
Not with 4dgspl.at. A single viewpoint does not contain enough information to rebuild a moving scene that can be watched from any angle. The service needs synchronized video of the same moment from at least 4 cameras; 8 are recommended.
Do I need my own GPU?
No. Processing runs on our GPU server; you only upload the videos and calibration in the browser or with the upload script.
What is the output format?
A browser player with a share link and iframe embed code, plus a ZIP with one 3D Gaussian Splatting file per frame (compressed PLY), sahne.json and an MP4 preview.
How long can the clip be?
The processing server currently uses the first 4 seconds of each capture at 15 fps (60 frames). Longer clips are planned.
I have no calibration. What now?
Create it with COLMAP from one frame per camera; the capture guide lists the exact commands. For virtual cameras the calibration is exported directly, see Blender render to 4DGS. Automatic calibration on the server is in development.