Research · Thread 01
3D Gaussian Splatting
Turning ordinary photographs of a real place into a 3D scene you can walk through in a browser.
Every brokerage in Dubai had Matterport, so every brokerage's tours looked identical, and the brief was to find something better. What had moved in the field was 3D Gaussian splatting: it stores a scanned room not as modelled surfaces but as millions of tiny translucent blobs of colour, each stretched and tinted to match what the cameras saw, and redraws them fast enough to walk through. Method papers report their trade-offs on clean benchmark scenes, but a mirror-walled Dubai interior on a phone's memory budget is a different question, and a benchmark is how it gets answered. Vantage is what came out the other end, and it ran in the BARNES Dubai property-search flow through 2026.
Findings
Radiance fields on a phone budget
NeRF's neural net, swapped for millions of explicit Gaussians. Closing the gap between the method paper and a real Safari-iOS budget. Vantage: the shipped answer, which ran on BARNES Dubai's property search through 2026.
Method-paper quality, running on a mobile budget.
Which engine actually deploys?
26 engines, one harness: studio, hard, and city-scale scans. PSNR/SSIM/LPIPS evaluation closed before the full matrix.
No ranked winner. The working hypothesis was that the quality winner would not be the speed winner.
Paper to shipped
- 3D Gaussian Splatting for Real-Time Radiance Field Rendering (Kerbl, Kopanas, Leimkühler, Drettakis, SIGGRAPH 2023)
The original 3DGS paper. Replaces NeRF's MLP with millions of explicit Gaussians, competitive training times and real-time (30fps+) novel-view synthesis.
View abstract
The authors introduce three elements that together achieve state-of-the-art visual quality while maintaining competitive training times and high-quality real-time (≥30 fps) novel-view synthesis at 1080p. Scenes are represented as 3D Gaussians initialised from sparse SfM points; an interleaved optimisation + adaptive-density-control step shapes the radiance field; a fast tile-based rasteriser keeps the render path GPU-resident. Demonstrated on several benchmarks against the prior NeRF state of the art.
- 26-engine 3DGS pipeline benchmark (evaluation closed)
26 Gaussian-splat engines wired into one harness across studio scans, hard scans, and city-scale capture. The evaluation closed before the full matrix. It used the three standard image-quality scores: PSNR for raw pixel error, SSIM for structural similarity, LPIPS for whether two images look alike to a human.
- Production-deployment lessons from the benchmark (not accepted)
Four pages, SIGGRAPH Asia 2026 Technical Communications format, anonymised for double-blind review and audited for citation and claim accuracy before being called finished. Submitted to SIGGRAPH 2026 and SIGGRAPH Asia 2026, and not accepted at either.
- Vantage in the BARNES property-search flow
Web-native viewer with tap-to-move navigation, LOD streaming, and golden-hour relighting. First georeferenced building (Onyx Tower) live for brokers.
