On this page
Gaussian Splatting
Gaussian Splatting
A real-time 3D scene representation technique introduced by Inria researchers in 2023. It reconstructs 3D environments from 2D photographs using millions of small semi-transparent ellipsoids ("splats"), each defined by position, orientation, scale, colour (via spherical harmonics), and opacity. Unlike NeRF (which requires ray marching), 3DGS renders via a specialised rasteriser — enabling 60+ FPS on consumer hardware where NeRF takes seconds per frame. As of 2026, the technique is entering ecommerce product visualisation, virtual try-on, and spatial showroom experiences via browser-based WebGPU renderers, a Shopify app ecosystem, and an emerging Khronos open standard.
How it works
- claim: 3DGS reconstructs 3D scenes by positioning millions of Gaussian distributions in 3D space and rendering them in real-time via a specialised rasteriser, not ray marching. (utsubo.com 2026-01-15)
- claim: 3DGS is significantly better than traditional photogrammetry for transparent and reflective objects — fashion jewellery and luxury goods that produced "erroneous and unusable surfaces" in mesh-based capture now work. (utsubo.com 2026-01-15)
- claim: The Khronos Group published a Release Candidate for the KHR_gaussian_splatting glTF 2.0 extension on 2026-02-03, establishing an open baseline for storing 3DGS assets in glTF — the most widely adopted 3D delivery format. The extension adds Gaussian splat datasets as a new mesh primitive type; renderers that don't support splatting fall back gracefully to a point cloud. Ratification was expected Q2 2026; a Bevy engine changelog entry (~2026-09-08) infers ratification occurred, but no Khronos ratification press release was confirmed in this run. (khronos.org PRIMARY 2026-02-03; bevyengine PR ~2026-09-08)
- claim: KHR_gaussian_splatting is algorithm-agnostic and extensible; the base spec uses an uncompressed ellipsoid kernel and is designed to accommodate future kernel types, colour spaces, projection methods, and sorting techniques. (khronos.org PRIMARY 2026-02-03)
- claim: Two vendor compression extensions were proposed alongside the base spec at RC time: Niantic Spatial's SPZ format (KhronosGroup PR #2531) and Qualcomm's L-GSC format (PR #2551). (khronos.org PRIMARY 2026-02-03)
- claim: Contributing organisations to KHR_gaussian_splatting include Autodesk, Cesium/Bentley Systems, Esri, Huawei, Niantic Spatial, NVIDIA, and XGRIDS. Early adopters committing support at RC time: CesiumJS, Esri ArcGIS, Niantic Scaniverse, XGRIDS Lixel Cybercolor. (khronos.org PRIMARY 2026-02-03)
- claim: Khronos and AOUSD (Alliance for OpenUSD) are actively coordinating to ensure glTF extensions and USD (Universal Scene Description)|OpenUSD schemas for Gaussian splats remain aligned, enabling seamless conversion between the two standards. (khronos.org PRIMARY 2026-02-03)
- claim: OGC (Open Geospatial Consortium) co-announced the addition of 3DGS to the glTF standard; KHR_gaussian_splatting is expected to become a component of the upcoming OGC 3D Tiles 2.0 standard. (ogc.org PRIMARY 2026)
Performance and file sizes
- claim: Raw 3DGS captures: 5–15 million splats at 200–500 MB; unoptimised, load times are unacceptable on web and mobile devices crash or freeze (as-of 2026-01-15). (utsubo.com 2026-01-15)
- claim: Web-optimised targets by device: Desktop 1–2M splats, 50–100 MB; Mobile 200–500K splats, 15–30 MB (as-of 2026-01-15). (utsubo.com 2026-01-15)
- claim: Typical FPS by device class after web optimisation (as-of 2026-01-15): Desktop RTX 3080 = 60+ FPS; M1 MacBook = 45–60 FPS; iPhone 14+ = 30–45 FPS. (utsubo.com 2026-01-15)
- claim: Compression using SPZ/SOG/KSPLAT formats yields 5–10× file size reduction vs raw PLY (as-of 2026-01-15). PlayCanvas open-sourced the SOG format, described as "the WebP of Gaussian Splatting." (utsubo.com 2026-01-15)
- claim: WebGPU delivers a 2–5× performance improvement over WebGL for 3DGS in typical browser scenarios; Chrome, Edge, Firefox, and Safari 26 all support WebGPU as of 2026. (utsubo.com 2026-01-15)
- claim: Switching the Reflct ecommerce viewer from the GaussianSplats3D three.js viewer to PlayCanvas nearly doubled frame rate and reduced memory usage by 80%; switching to PlayCanvas SOG format halved file bundle size and reduced memory by 88%, with some resources consuming 95% less memory (as-of 2025-08-21).
- claim: Mobile-GS research (OpenReview 2025) achieved 4.8 MB splat models rendering at 1,098 FPS on mobile via depth-aware rendering + distillation. As of 2026 this is a research result, not a generally available shipping standard. (openreview.net/forum?id=vRegY0pgvQ 2025)
Ecommerce tools and platforms
- claim: Reflct (reflct.app, NZ-based) is a purpose-built 3DGS ecommerce viewer on PlayCanvas; its Shopify app lets brands capture a showroom or staged space and attach product metadata to each viewpoint, creating a "shoppable" 3DGS environment where product details populate a side panel as the shopper moves between products. Dashboard auto-compresses uploaded PLY files; integration is via React library or Shopify component. (blog.playcanvas.com 2025-08-21)
- claim: 3D Bits (by Bitbybit/bitbybit.dev) is a Shopify app ($142/yr) that embeds .splat Gaussian Splatting files directly on Shopify product pages via a BITBYBIT VIEWER block; files can be loaded from Shopify's CDN or any public URL. Recommended capture workflow: scan via Scaniverse (iOS/Android) → export .splat → upload to Shopify CDN → configure in Bitbybit node graph → inject into product page. (bitbybit.dev; YouTube 2024-07-31)
- claim: Spark (by World Labs, built on Three.js) was named one of GitHub's most influential libraries of 2025 and is described by practitioners as the current leader for production 3DGS on websites (as-of 2026-01-15). (utsubo.com 2026-01-15)
- claim: RaveSpace (Berlin) integrates 3DGS via its proprietary RAVE ENGINE to generate photorealistic 3D assets from client-provided videos/images; use cases include virtual showrooms with multiplayer and voice chat. (ravespace.io 2025-09-10)
- claim: WEARFITS states 30+ standard studio photographs are sufficient to generate a functional 3D model via their AI converter, eliminating the need for CAD files or multi-camera arrays. (wearfits.com 2026-03-26; vendor claim, unverified independently)
- claim: No Khronos 3D Commerce Working Group guidance specific to Gaussian Splatting for retail product visualisation was found as of 2026-09-22. The connection between KHR_gaussian_splatting and retail use cases exists only at general "e-commerce" level in analyst commentary, not in official 3D Commerce Working Group documentation. (registry-source, this run)
Fashion, virtual try-on, and 4DGS
- claim: Gracia (London startup) demonstrated 4D Gaussian Splatting (4DGS) — animated volumetric video — streaming directly to VR headsets and standard web browsers; it secured $1.7M in funding and delivered the first 4DGS runway experience for Karl Kani. Compression achieves near-lossless reduction to files starting at 1.5 GB per minute of animated content; supports Meta Quest 3/3S, Pico 4 Ultra, and WebGPU-based Mac/browser viewing. (wearfits.com 2026-03-26; via Crunchbase)
- claim: GS-VTON (arXiv Oct 2024) demonstrated controllable 3D virtual try-on using Gaussian Splatting. Not yet commercialised as of the run date. (arxiv.org/pdf/2410.05259 2024-10)
- claim: A dedicated volumetric capture studio for 4DGS fashion try-on costs $50,000–$150,000 to build (synchronised cameras, lighting rigs, processing servers); many retailers hire external agencies (as-of 2026-03-26). (wearfits.com 2026-03-26)
- claim: Fashion ecommerce return rates run approximately 24–26% for apparel and 31.4% for footwear — the highest of any product category; poor fit accounts for 38% of online apparel returns (as-of 2025). (wearfits.com 2026-03-26; Capital One Shopping 2025)
Industry adoption signals
- claim: Apple released SHARP in early 2026, a technique that generates high-quality 3DGS from a single image — implying product catalogue photos could be converted to 3D assets at scale. (utsubo.com 2026-01-15; no primary Apple source confirmed)
- claim: Zillow became the first major platform to ship Gaussian Splatting via SkyTours (real estate exterior tours); CoStar Group / Apartments.com followed via Matterport 3D Exteriors. These are the primary confirmed production deployments — ecommerce product-level named case studies remain sparse. (utsubo.com 2026-01-15; Volinga 2026-02-20)
- claim: The production of the film Superman was the first major motion picture to ship with dynamic Gaussian Splatting, using Houdini-based 3DGS tools inside a professional VFX pipeline — cited as a signal that the technology is production-ready. (web.volinga.ai 2026-02-20)
- claim: The Foundry announced native 3D Gaussian Splat support in Nuke 17.0 open beta in late 2025. (web.volinga.ai 2026-02-20)
- claim: Cesium implemented KHR_gaussian_splatting support in CesiumJS and Unreal Engine following the RC publication. (radiancefields.com 2026)
- claim: Khronos hosted "Gaussian Splats: Use Cases & Next Steps for glTF Standardization" webinar on 2026-04-07 featuring Cesium, XGRIDS, and the Smithsonian. (khronos.org PRIMARY 2026-04-07)
- claim: KIRI Engine CEO and RadianceFields.com founder Michael Rubloff described 2025 as the year 3DGS moved "from research curiosity to production reality" and 2026 as the inflection point where it becomes a standard tool for marketing, advertising, and ecommerce. (YouTube/KIRI Engine, published 2025-12-25; RadianceFields substack 2025-12)
Benchmarks and market projections
- claim: DIY capture (smartphone + Polycam/Luma AI) costs free to ~$50/month; professional capture $500–$5,000 depending on space size; custom web viewer development $5,000–$50,000+; enterprise multi-property systems priced custom (as-of 2026-01-15). (utsubo.com 2026-01-15; vendor pricing guide)
- claim: Grand View Research projects the virtual try-on machine-learning segment will grow at 30.1% CAGR through 2030 (as-of 2026-03-26). (Grand View Research cited in wearfits.com 2026-03-26)
- claim: Fortune Business Insights projects the global volumetric video market will reach $35.03 billion by 2034 (as-of 2026-03-26). (Fortune Business Insights cited in wearfits.com 2026-03-26)
- claim: Only 7% of companies have moved beyond AI experimentation to fully scaled deployments; 62% remain in piloting phases (as-of 2026) (Envive 2026 generative AI commerce analysis, cited in wearfits.com 2026-03-26).
Contradictions
Key terms
| Term | Meaning |
|---|---|
| 3DGS / Gaussian Splatting | 3D scene representation using millions of Gaussian ellipsoids rendered in real-time via rasteriser |
| 4DGS | Animated/temporal Gaussian Splatting; volumetric video extending 3DGS with a time dimension |
| PLY | Common raw Gaussian Splatting file format (large, uncompressed) |
| SPZ | Niantic Spatial's compressed Gaussian Splatting format |
| SOG | PlayCanvas open-source compressed format; described as "WebP of 3DGS" |
| KSPLAT | Another compressed splat format |
| KHR_gaussian_splatting | Khronos glTF 2.0 extension adding 3DGS as a native mesh primitive type |
| Spherical harmonics | Mathematical representation used to encode colour as a function of viewing angle in 3DGS |
What practitioners report
- Primary deployment barrier is initial payload size and compression pipeline, not real-time rendering performance once loaded (utsubo.com 2026; PlayCanvas/Reflct 2025)
- Viewer choice driven by library maintenance trajectory as much as render quality — Reflct switched from mkkellogg to PlayCanvas as mkkellogg support faded (PlayCanvas blog 2025)
- Shopify integration is currently a technical merchant workflow (5 steps via 3D Bits at $142/yr) not a self-serve tool for average merchants as of 2026 (RadianceFields.com 2024; Bitbybit tutorial 2024)
- 3DGS handles transparent and reflective materials better than mesh photogrammetry — identified as key advantage for jewellery, luxury, and accessories (utsubo.com 2026; multiple practitioner sources)