Every tool and workflow used to take drone footage of a Dartmoor heritage site and turn it into a live, navigable 3D Gaussian splat — from capture in the field through to gsplat.js running in the browser.
Quality in, quality out — the capture stage determines everything downstream. For aerial heritage survey the DJI Mini 4 Pro delivers excellent overlap at low altitude. Luma AI is the fastest path for quick-turnaround scenes captured on foot.
Structure-from-Motion produces the camera positions and sparse point cloud that 3DGS training builds from. COLMAP is the open-source standard. Postshot wraps SfM + training in a single GUI for local processing without touching the command line.
Raw trained splats always contain floaters — stray Gaussians from motion blur, sky reconstruction, or insufficient coverage. Cleanup before deployment is essential, and often reduces file size by 70–90% with no visible quality loss.
The goal for Hylas Spatial is a cleaned .splat file rendering live in a browser with no plugin, no install, and no server-side compute — just static files and WebGL.
| Tool | Local | GUI | SfM | Trains 3DGS | Edits splats | Web deploy | Cost |
|---|---|---|---|---|---|---|---|
| Luma AI | Cloud | ✓ | ✓ | ✓ | Basic | ✓ | Free tier |
| Postshot | ✓ | ✓ | ✓ | ✓ | Basic | — | Free (beta) |
| COLMAP | ✓ | CLI/GUI | ✓ | — | — | — | Free |
| Inria 3DGS | ✓ | CLI | — | ✓ | — | — | Free |
| Nerfstudio | ✓ | CLI | ✓ | ✓ | — | — | Free |
| SuperSplat | Browser | ✓ | — | — | ✓ | ✓ | Free |
| SplatForge | ✓ | ✓ Blender | — | — | ✓ | — | $49 |
| gsplat.js | ✓ | Code | — | — | — | ✓ | Free |
The active Hylas pipeline runs: Waypoint OS → DJI Fly → ffmpeg → COLMAP → 3DGS trainer → SuperSplat → gsplat.js. For quick scene tests, Luma AI handles the cloud path. Postshot is used for local training of complex scenes where COLMAP alignment needs visual inspection. Cleaned splats are served as static files on IONOS — no backend, no database, no server compute.