Technology Comparison

3D Gaussian Splatting vs NeRF: Neural Rendering Methods Compared

An expert comparison to help you choose the right equipment for your project.

Feature	3D Gaussian Splatting	NeRF (Neural Radiance Fields)
Training Time	2-7 minutes (vs 18+ hours for NeRF)	18-20 hours (original)
Rendering Speed	100+ FPS real-time	~5 FPS historical; 30 FPS optimized
Memory Usage	1-2 GB typical scene	200-500 MB (compact)
Quality (PSNR)	25-33 dB (higher than NeRF)	25-31 dB
Editability	Explicit primitives — easy to edit	Implicit — difficult to modify
Output Formats	PLY, 3DTiles, glTF, OpenUSD	-
Professional Tools	DJI Terra, Polycam, Luma AI, PostShot	-
Editing Tools	SuperSplat, SplatForge (Blender)	-
Mean Geometric Error	7.82 cm ± 11.49 cm (not survey-grade)	-
Service Pricing	Professional GS processing service: 1.5× photogrammetry rates (minimum $2,250). DJI Terra Flagship license: $2,800-$4,400. DIY: Polycam, Luma AI, PostShot available free or low-cost.	Open-source via Nerfstudio; Research-focused — no commercial service providers

Pricing shown reflects average US rates. Actual costs vary by location based on local market conditions, regulations, and project logistics — both within the US and internationally. Get a custom quote

Option A Neural Rendering (Inria, SIGGRAPH 2023)

3D Gaussian Splatting

Real-Time Photorealistic Rendering at 100+ FPS

3D Gaussian Splatting (3DGS), introduced by Inria at SIGGRAPH 2023, is a breakthrough neural rendering technique that represents 3D scenes as collections of anisotropic Gaussian ellipsoids. Unlike NeRF, 3DGS uses explicit geometric primitives that can be rasterized directly by the GPU — no neural network inference is required during rendering. This enables real-time photorealistic visualization at over 100 FPS on consumer hardware. Professional tools now support 3DGS natively: DJI Terra V5.0+ processes drone imagery into Gaussian Splats, Polycam and Luma AI offer mobile capture-to-GS pipelines, and editing tools like SuperSplat and SplatForge enable post-processing of splat scenes. The technology has been adopted into industry standards including OpenUSD (April 2026) and Khronos glTF (KHR_gaussian_splatting extension).

Pros

Real-time rendering at 100+ FPS on consumer GPUs
Training completes in minutes, not hours
Explicit representation allows direct editing and manipulation
Higher visual quality (25-33 dB PSNR vs 25-31 dB for NeRF)
GPU-efficient tile-based rasterization pipeline
DJI Terra V5.0+ native support for professional drone workflows
Growing ecosystem: Polycam, Luma AI, PostShot, SuperSplat, SplatForge
Adopted into OpenUSD and Khronos glTF industry standards

Cons

Higher memory requirements than mesh or NeRF representations
Viewer compatibility still evolving (improving rapidly)
Not suitable for engineering measurements (7.82 cm mean geometric error)
Larger file sizes than compact NeRF models
CAD/BIM software integration still limited

Best For

Virtual tours and walkthroughsReal estate and property marketingConstruction site visualizationFilm and virtual production environmentsHeritage site documentationUrban planning visualization

Option B Neural Rendering (UC Berkeley, 2020)

NeRF (Neural Radiance Fields)

The Pioneering Neural Scene Representation

Neural Radiance Fields (NeRF), introduced by UC Berkeley researchers in 2020, represents 3D scenes as continuous volumetric functions learned by a neural network. For each point in 3D space, a trained neural network predicts color and density, which are then volumetrically rendered into novel views. NeRF produces high-quality novel view synthesis but requires 18-20 hours of training and cannot render in real-time on standard hardware without significant optimization. Research tools like Nerfstudio provide an open-source framework for training and evaluating NeRF variants. While 3DGS has largely superseded NeRF for production applications, NeRF remains a foundational technology in computer vision research and has inspired numerous variants including Instant-NGP, Mip-NeRF 360, and TensoRF.

Pros

Compact representation requires lower storage
Foundational technology with extensive research literature
Continuous volumetric representation handles thin structures
Nerfstudio provides robust open-source training framework
Works well for view interpolation and synthesis tasks

Cons

Very long training times (18+ hours for original NeRF)
Not real-time on standard hardware without optimization
Implicit representation is difficult to edit or manipulate
Requires neural network inference for every rendered pixel
Visual quality limited compared to 3DGS on most benchmarks
No major professional production tool supports NeRF natively

Best For

Academic and computer vision researchVR/AR view synthesis experimentsCompact scene storage where file size is criticalExisting Nerfstudio-based workflows

Our Expert Verdict

Winner: 3D Gaussian Splatting

3D Gaussian Splatting has largely superseded NeRF for practical applications. 3DGS renders at 100+ FPS versus NeRF's ~5 FPS, trains in minutes instead of hours, and achieves equal or higher visual quality (25-33 dB PSNR). Professional tools like DJI Terra V5.0+, Polycam, and Luma AI now support GS natively, and industry standards OpenUSD and glTF have adopted the format. NeRF remains valuable for academic research and situations where compact file size is the primary concern.

Choose 3D Gaussian Splatting if...

Choose 3D Gaussian Splatting for production work: virtual tours, construction visualization, real estate marketing, film/VFX environments, heritage documentation, and any application requiring real-time interaction. Professional GS services are available from scanning companies like THE FUTURE 3D.

Choose NeRF (Neural Radiance Fields) if...

Choose NeRF for academic research, computer vision experimentation, situations where compact file size outweighs rendering speed, or when working within existing Nerfstudio-based pipelines.

Frequently Asked Questions

What is 3D Gaussian Splatting?

3D Gaussian Splatting (3DGS) is a neural rendering technique introduced by Inria at SIGGRAPH 2023. It represents 3D scenes as collections of Gaussian ellipsoids — each with position, orientation, opacity, and color properties. These Gaussians are "splatted" (projected and blended) onto the screen using GPU rasterization, enabling real-time photorealistic visualization at 100+ FPS without requiring neural network inference during rendering.

Why is 3DGS faster than NeRF?

NeRF requires running a neural network for every pixel during rendering — a computationally expensive operation. 3DGS uses explicit Gaussian primitives that can be rendered using standard GPU rasterization pipelines, a much faster operation that enables 100+ FPS on consumer hardware like an NVIDIA RTX 4060. This speed advantage makes 3DGS practical for real-time applications like virtual tours and interactive visualization.

What is the accuracy of Gaussian Splatting?

Gaussian Splatting achieves a mean geometric error of 7.82 cm ± 11.49 cm standard deviation, based on independent research (plainconcepts.com). This means GS produces sub-centimeter visual fidelity but is NOT suitable for engineering measurements or survey-grade work. For projects requiring precise measurements, professional scanning companies like THE FUTURE 3D combine GS photorealism with ±1-2mm LiDAR accuracy from instruments like the Trimble X12.

Does DJI Terra support Gaussian Splatting?

Yes. DJI Terra V5.0+ (Flagship, Cluster, and Education licenses only — not Standard) supports 3D Gaussian Splatting processing. It converts drone aerial imagery into GS scenes at approximately 500 images per hour, handles up to 30,000 photos, and outputs 3DTiles (for Cesium web viewers), PLY (Gaussian Splats), and GeoTIFF formats. GPU requirements: NVIDIA RTX with 8GB+ VRAM, 32GB+ system RAM recommended.

What other tools support Gaussian Splatting?

Beyond DJI Terra, several tools support GS: Polycam (iOS/Android/Web — mobile capture to GS), Luma AI (iOS/Web — free cloud processing), PostShot (Windows — high-quality desktop processing), Xgrids LCC (LiDAR-to-GS with Revit plugin), and 3DMakerPro RayStudio (spatial scanner to GS in standard PLY). For editing, SuperSplat (web-based, open-source) and SplatForge (Blender add-on supporting 16M+ splats) are the primary tools.

Can you combine Gaussian Splatting with LiDAR?

Yes. Combining GS with LiDAR is the most powerful approach for professional applications. THE FUTURE 3D uses this hybrid method: survey-grade LiDAR scanners like the Trimble X12 (±2mm accuracy) capture precise measurements, while GS from DJI Terra adds photorealistic visualization. This delivers both engineering-grade accuracy AND visual fidelity — something no competitor currently matches. Xgrids LCC also offers LiDAR-to-GS processing with a Revit plugin for BIM workflows.

What file formats does Gaussian Splatting use?

GS scenes are primarily stored in PLY format (containing Gaussian parameters: position, covariance, opacity, spherical harmonics for color). DJI Terra also outputs 3DTiles for web-based LOD streaming via Cesium. Industry standards are rapidly adopting GS: OpenUSD added official GS support in April 2026, and Khronos is developing the KHR_gaussian_splatting glTF extension. The SOG (Streamed Oriented Gaussians) format enables level-of-detail streaming for 10M+ splat scenes.

Can I use 3DGS for engineering measurements?

No. 3D Gaussian Splatting is optimized for photorealistic visualization, not measurement. With a mean geometric error of 7.82 cm, GS cannot meet engineering tolerances. For measurement-critical work requiring ±1-4mm accuracy, use point clouds from LiDAR or survey-grade photogrammetry. THE FUTURE 3D delivers both: GS for visualization and point clouds or orthomosaics for measurement — as separate, complementary deliverables.

Do you scan individual objects with Gaussian Splatting?

No. THE FUTURE 3D specializes in scanning buildings, environments, locations, and sites — not individual objects, props, or products. Our GS services cover building exteriors and interiors, construction sites, film/TV locations, heritage sites, and urban environments. For object-level scanning, consumer tools like 3DMakerPro Eagle or Polycam are better suited.

When should I still use NeRF instead of GS?

NeRF may still be preferred for academic research and computer vision experimentation (Nerfstudio provides robust tooling), situations where compact file size is critical (NeRF models are 200-500 MB vs 1-2 GB for GS), or when working within existing NeRF-based pipelines. For all production and commercial applications — virtual tours, AEC visualization, film/VFX, real estate marketing — 3DGS is now the better choice due to real-time rendering, faster training, and growing professional tool support.

Which does THE FUTURE 3D use?

THE FUTURE 3D uses 3D Gaussian Splatting via DJI Terra V5.0+ for visualization deliverables — virtual tours, marketing content, client presentations, and film/VFX environments. We process drone imagery from our DJI M4E fleet through Terra's GS pipeline, outputting 3DTiles for web viewers and PLY for downstream editing. For measurement-critical deliverables, we provide point clouds and survey-grade orthomosaics from our Trimble X12 (±2mm) and NavVis VLX 3 (±5mm) scanners. GS processing is a premium add-on service priced at 1.5× standard photogrammetry rates, with a minimum project cost of $2,250.

What is OpenUSD Gaussian Splatting support?

OpenUSD (Universal Scene Description), the industry-standard scene format developed by Pixar and maintained by the Alliance for OpenUSD, officially added Gaussian Splatting support in April 2026. This means GS scenes can now be integrated into professional VFX, animation, and virtual production pipelines that use USD — including Unreal Engine, NVIDIA Omniverse, and SideFX Houdini. Combined with the Khronos glTF extension (KHR_gaussian_splatting), GS is rapidly becoming an interoperable standard across the 3D industry.

Our Services