Amazon

Amazon Web Services publishes the Open Source 3D Reconstruction Toolbox for Gaussian Splats, a deployable guidance repository in the AWS Solutions Library that turns uploaded video or photos into a trained Gaussian splat inside the customer's own AWS account. The repository dates to March 2025 and tagged its first release, v1.0.0, on July 21, 2026. It is written mainly in Python and licensed under MIT No Attribution.

A job starts when media and a JSON job configuration land in an Amazon S3 bucket. An Amazon SNS message invokes an initialization AWS Lambda function, an AWS Step Functions state machine runs the workflow, and the reconstruction itself executes in a GPU container as an AWS Batch job on Spot capacity or as an Amazon SageMaker training job on demand. Amazon DynamoDB records each job, Amazon CloudWatch carries the training logs, and Amazon SNS emails the user when the asset is ready. Jobs can be submitted through a bundled Gradio web interface or written straight to the bucket, and the infrastructure deploys with Terraform or the AWS CDK.

Camera poses come from COLMAP or GLOMAP, with multi GPU parallel structure from motion, equirectangular to perspective conversion for 360 footage, support for images supplied with pose priors, and densification of sparse ARKit/LiDAR depth maps. Training runs on Nerfstudio and gsplat, and v1.0.0 lists splatfacto, splatfacto-big, splatfacto-mcmc, dn-splatter, ags-mesh, 3DGUT, 3DGRT, gsplat-depth and nerfacto among its models. Preprocessing covers frame extraction, blurry image filtering and background removal, and the v1.0.0 container ships CUDA 13, PyTorch 2.9 and COLMAP 4.0.4.

Exports cover PLY, SPZ and SOG splats, USDZ in beta, GLB meshes extracted from the splat, collision meshes for physics scenes and rendered video. PlayCanvas splat-transform handles format conversion and builds streamed multi level of detail SOG bundles, and the bundled viewer uses SuperSplat and Babylon.js.

AWS Batch job definitions cover NVIDIA A10G, L4 and L40S GPUs through the g5, g6 and g6e instance families, with ml.g5.4xlarge as the default. The toolbox itself is free. AWS estimates the default deployment in US East (N. Virginia) at about $73.33 a month for 100 one hour jobs on Spot instances, or about $142.33 a month using SageMaker on demand, figures it dates to May 2025.