Backblaze has published an open source template that turns synchronized multi-camera video into a trained 4D Gaussian splatting model stored in Backblaze B2.
Dynamic splats have mostly arrived as finished content, from Gracia streaming its 4DGS captures to Apple Vision Pro to an 11.5 minute video made entirely of 4DGS with 4DV.ai and Dell Technologies. The Backblaze project, 4D Gaussian Splatting Volumetric Capture, was published through Backblaze Labs in August and covers the stage before playback. It stages a dataset in the multipleview layout used by hustvl/4DGaussians, the code behind the CVPR 2024 4D Gaussian Splatting paper, which models motion as a deformation field over a set of gaussians.
Ingest, frame extraction with a bundled ffmpeg, calibration, a sparse initial point cloud, dataset staging, previews and all storage reads and writes run on CPU. A setup script clones 4DGaussians and builds its two CUDA submodules, after which the app calls the engine's train.py in an isolated subprocess and uploads checkpoints and the final .ply splat to B2. On a host without CUDA the training stage is gated off and the app prints the exact train.py command to run on a GPU machine.
Backblaze builds the project around write amplification, the way one capture session fans out from source video into thousands of extracted frames, calibration data, multi gigabyte checkpoints and a final splat. A dashboard charts that fan out per session, and a storage explorer breaks each session down by pipeline stage. Each session is a JSON manifest in the bucket with no database behind it, accessed over the S3 compatible API with boto3.
The web app runs on Next.js 16 and React 19, and the API on FastAPI and Pydantic v2. A seed script renders a synthetic multi-camera capture and runs the CPU pipeline end to end. The template deploys to Vercel as a single project, where training does not run.
Backblaze labels the project a sample, unsupported and without an SLA, and the README describes production use as "possible with caution". The deployed API is unauthenticated with bucket wide access, so Backblaze recommends a dedicated bucket or prefix and key for any preview.
The template is available now on GitHub under the MIT license.
