3D Gaussian Splatting: Services, Use Cases, and Web Viewers (2026)

Research Date: 2026-02-06 Source URL: https://www.utsubo.com/blog/gaussian-splatting-guide Author: Jocelyn Lecamus, Co-Founder/CEO of Utsubo Published: 2026-01-15 Shared via: Bert Temme (@berttemme) on 2026-02-04

Reference URLs

• Utsubo Gaussian Splatting Guide
• Original 3DGS Paper - Kerbl et al., SIGGRAPH 2023
• Spark Viewer (sparkjs.dev)
• Polycam
• Luma AI
• Postshot
• GaussianSplats3D (GitHub)
• antimatter15/splat (GitHub)
• Utsubo Portfolio

Summary

This article by Jocelyn Lecamus of Utsubo provides a market-oriented survey of 3D Gaussian Splatting (3DGS) as of early 2026, covering the full pipeline from scene capture through web deployment. The guide distinguishes itself from academic or purely technical treatments by mapping the commercial ecosystem: which services handle capture, which tools process and compress splat data, which viewers deliver the final experience on the web, and what each option costs.

The article positions 3DGS as having crossed the threshold from research prototype to production-ready technology. Real-time rendering at 60+ FPS on consumer hardware, combined with training times measured in minutes rather than hours, makes 3DGS viable for real estate, e-commerce, film production, cultural heritage, and education. The analysis identifies web delivery as the primary remaining friction point, with raw captures (5-15 million splats, 200-500 MB) requiring aggressive optimization to reach acceptable load times on desktop and mobile browsers.

Key commercial signals include Zillow’s SkyTours integration, the Superman film production’s use of dynamic 4D Gaussian scenes, and broad game engine support through Unreal Engine plugins. Browser-side, full WebGPU support across Chrome, Edge, Firefox, and Safari (as of Safari 26 in 2025) removes the last major rendering bottleneck for web-based splat viewers.

Technology Overview

How 3D Gaussian Splatting Works

3DGS reconstructs 3D scenes from a set of photographs by representing them as millions of small, semi-transparent ellipsoids (3D Gaussians) rather than polygon meshes. The original paper by Kerbl, Kopanas, Leimkühler, and Drettakis (SIGGRAPH 2023) introduced three core innovations:

1. A 3D Gaussian representation initialized from sparse camera calibration points, preserving volumetric radiance field properties while skipping empty space
2. An interleaved optimization and density control scheme with anisotropic covariance refinement
3. A visibility-aware splatting renderer that achieves 100+ FPS at 1080p for complete, unbounded scenes

The training pipeline takes multi-angle photographs as input, performs Structure from Motion (SfM) to establish camera positions and a sparse point cloud, then optimizes Gaussian parameters (position, covariance, color, opacity) to minimize the difference between rendered and ground-truth images.

Comparison with Alternative 3D Capture Methods

The article provides side-by-side comparisons against the main competing approaches. These tables consolidate the article’s assessment.

3DGS vs. NeRF vs. Photogrammetry

3DGS vs. Matterport vs. 360 Photos

The primary advantage of 3DGS over Matterport is continuous navigation: users can move freely through a scene rather than teleporting between fixed scan points. The primary disadvantage relative to photogrammetry is editability: 3DGS scenes cannot be geometrically modified after capture without re-running the training pipeline.

Capture and Processing Services

Consumer and Prosumer Tools

Polycam targets the mobile capture workflow and integrates with iPhone LiDAR for depth-assisted reconstruction. Luma AI reduces friction to near zero by accepting video uploads and processing server-side. Postshot occupies the middle ground, offering desktop-based processing with more user control over output quality parameters.

Professional Services

For spaces too large or complex for consumer tools, the article identifies a professional capture tier at $500-$5,000+ per project. This range covers custom studio setups with controlled lighting, multi-camera rigs, and manual quality assurance. The upper end applies to large commercial or architectural spaces requiring high fidelity.

Open-Source Processing Tools

The INRIA implementation serves as the reference standard but requires CUDA-capable hardware and comfort with Python environments. Nerfstudio wraps multiple Gaussian splatting backends into a unified developer framework.

Web Viewers

The viewer layer translates trained 3DGS scenes into interactive web experiences. The article evaluates five options:

Spark Viewer

The article identifies Spark as the current leading open-source option for production deployments. Spark is built on Three.js and supports the widest range of file formats. Its core architecture includes:

• SparkRenderer: Main rendering pipeline with WebGPU support in development
• SplatMesh: Integration primitive for mixing splats with conventional Three.js meshes
• PackedSplats: Compressed splat data handling
• Dyno: Programmable dynamic splat effects system

Spark is actively developed and offers Discord community support alongside GitHub issue tracking. The article notes that Spark is developed by World Labs, though the Spark website itself does not confirm this relationship directly.

Viewer Selection Criteria

Web Optimization Pipeline

Raw 3DGS captures are not web-ready. A scene captured at full quality typically contains 5-15 million splats and occupies 200-500 MB on disk. The article outlines a multi-stage optimization pipeline to bring these numbers within acceptable bounds for web delivery.

Optimization Stages

Target Performance by Device Class

Mobile support requires iPhone 12 or later, or recent Android flagships. The article recommends CDN deployment for all tiers given the 20-200 MB initial load sizes.

File Formats

The article recommends maintaining PLY for archival purposes while using compressed formats (SPZ, KSPLAT, or SOG) for web delivery. The existing note on SOG format and PlayCanvas streaming in this repository covers SOG’s technical architecture in depth, including its WebP-based encoding scheme and codebook quantization.

WebGPU Readiness

WebGPU replaces WebGL as the low-level graphics API for browsers, offering 2-5x performance improvements through compute shaders and better memory management. The article reports the following browser support status as of early 2026:

With all major browsers now supporting WebGPU, the article positions it as the target rendering API for new viewer implementations. Spark is noted as having WebGPU development underway with an auto-fallback to WebGL for older browsers.

Industry Use Cases and Adoption

Real Estate

The article highlights real estate as the most commercially advanced 3DGS use case. Zillow’s SkyTours product and Apartments.com (CoStar Group) have integrated 3DGS-based virtual tours. The smooth continuous navigation advantage over Matterport’s point-to-point teleportation model is the primary selling point for property marketing.

Applications include:

• Residential property virtual tours
• Pre-construction architectural visualization
• Renovation preview rendering
• Commercial real estate investor presentations and pre-leasing

Film and Entertainment

The Superman production is cited as the first major motion picture to use dynamic 3D Gaussian Splatting for virtual production environments. Broader applications span:

• Virtual production LED stage environments
• VFX asset creation from real-world scans
• 4D performance capture (time-varying Gaussians)
• Documentary and historical scene reconstruction

E-commerce and Retail

3DGS excels at capturing material properties that traditional 3D scanning methods struggle with: fabric texture, jewelry reflections, wood grain. The article identifies furniture, fashion, luxury goods, and jewelry as the strongest product categories for 3DGS-based product visualization. AR furniture placement is noted as an adjacent application.

Cultural Heritage and Museums

3DGS captures surface qualities like patina, weathering, and material aging that mesh-based digitization tends to lose. Use cases include artifact preservation, virtual museum tours, archaeological reconstruction, and pre-restoration documentation of damaged sites.

Education and Training

Virtual lab exploration, historical space reconstruction, medical specimen examination, and technical machinery inspection are identified as education-sector applications. The barrier is primarily awareness and integration effort rather than technical limitation.

Adoption Signal Summary

Pricing and Timeline

Budget Tiers

Project Timelines

Future Directions

The article identifies four active development vectors:

1. 4D Gaussian Splatting: Extending 3DGS with a time dimension for capturing motion, performances, and dynamic environments
2. AI-generated 3DGS: Apple’s SHARP model generates 3DGS assets from single input images, reducing the multi-view capture requirement
3. Native browser support: Early-stage discussions on integrating splat rendering primitives into web standards, which would eliminate the need for JavaScript viewer libraries
4. WebXR integration: Tighter coupling between 3DGS viewers and the WebXR protocol for VR and AR headset delivery

Limitations Noted in the Article

Editing constraints: 3DGS scenes allow only minor post-capture adjustments (splat removal, color correction). Geometric modifications require re-capture and re-training.

Web delivery friction: The optimization pipeline from raw capture to web-ready asset remains a multi-step process requiring specialized knowledge.

Commercial positioning: The article is published by Utsubo, a studio that sells 3DGS web integration services. While the technical content is accurate and well-sourced, the service comparisons and recommendations should be read with this commercial context in mind. Utsubo specializes in Three.js development and WebGPU-ready Gaussian splatting viewers.

References

1. Utsubo Gaussian Splatting Guide - Published 2026-01-15, accessed 2026-02-06
2. 3D Gaussian Splatting for Real-Time Radiance Field Rendering - Kerbl et al. - SIGGRAPH 2023
3. Spark Viewer Documentation - Accessed 2026-02-06
4. Bert Temme tweet sharing the article - 2026-02-04