3D Gaussian Splatting: The Rendering Revolution Killing Traditional Polygons

For decades, the 3D graphics industry has been obsessed with triangles. If you wanted to digitize a real-world object or environment, you relied on photogrammetry: taking hundreds of photos and forcing software to guess how to stretch a 2D texture over a jagged, million-polygon mesh. It was slow, computationally brutal, and fundamentally flawed. If the object was shiny, transparent, or lacked distinct texture, your final render inevitably looked like a melted candle.

Before you read: We recently explored how AI is hallucinating playable worlds frame-by-frame in Neural Game Engines: The End of the Traditional Render Pipeline. While those models throw out geometry to predict pixels, Gaussian Splatting throws out polygons to paint with millions of mathematical fuzzy blobs.

Table of contents

• The Problem with NeRFs and Polygons
• Enter the Splat
• The FPS Leap and the Memory Catch
• Fun Facts
• The End of the Mesh
• Sources

The Problem with NeRFs and Polygons

To understand why the industry is currently losing its mind over Gaussian Splatting, you have to look at what came immediately before it. In 2020, Neural Radiance Fields (NeRFs) blew everyone away. Instead of building a 3D mesh, NeRFs used a neural network to calculate how light rays traveled through a volume based on 2D photos.

It was a massive leap in photorealism, but it had a fatal flaw: speed. Asking a neural network to calculate the color and density of every single pixel along a light ray for every frame is a computational nightmare. NeRFs struggled to hit 10 frames per second on high-end hardware. It was a beautiful science experiment that you couldn’t actually play a game in.

Traditional photogrammetry is like trying to build a Lego model of a cloud. NeRFs are like asking an AI to paint the cloud from memory every single frame. Gaussian Splatting is like throwing millions of fuzzy paintballs into the air and telling them exactly where to freeze.

Enter the Splat

Introduced in a landmark 2023 paper, 3D Gaussian Splatting abandons neural networks during the actual rendering phase. Instead, the algorithm analyzes your 2D photos and generates millions of 3D ellipsoids—essentially, mathematical blobs. Each blob has a specific position, color, opacity, and rotation.

To render the scene, the GPU just sorts these blobs from back to front and squashes them flat against your 2D screen. GPUs absolutely love this. It is a highly optimized rasterization process that completely bypasses the complex ray-marching math of NeRFs. You get the photorealism of a volumetric scene, capturing flawless reflections and transparencies, running at well over 100 FPS at 1080p resolution.

I honestly can’t remember the last time a rendering technique disrupted the industry this fast. Researchers named it “splatting,” which sounds exactly like a sound effect from a 1990s Nickelodeon slime show, but it is currently the most feared word for anyone selling enterprise photogrammetry software.

The FPS Leap and the Memory Catch

So, if Gaussian Splatting is incredibly fast and looks indistinguishable from reality, why hasn’t Unreal Engine entirely replaced polygons yet?

Because splats are absolute memory hogs.

When you build a wall in a traditional video game, you use two triangles and a repeating texture file. That takes up a few kilobytes of VRAM. To represent that same wall using Gaussian Splatting, you need hundreds of thousands of individual blobs, each carrying its own mathematical parameters. A single, small room captured via splatting can easily balloon to 2GB of point cloud data.

We traded a compute bottleneck for a storage bottleneck. Streaming a 50GB multiplayer map made of polygons is standard practice today. Streaming a 500GB map made of billions of uncompressed mathematical splats will instantly melt your router.

Fun Facts

• The original 2023 paper by Inria and the Max Planck Institute won Best Paper at SIGGRAPH, practically overnight turning millions of dollars of enterprise photogrammetry software into legacy code.
• Because splats are just math equations representing volume, you can slice through them in real-time without the “hollow shell” effect you get when the camera clips through a traditional polygon wall in a video game.

The End of the Mesh

Engineers are already aggressively attacking the memory problem. Recent papers have demonstrated compression techniques that reduce splat file sizes by up to 30x without losing visual fidelity. We are watching the real-time death of the polygon mesh for real-world digitization.

Triangles will stick around for purely synthetic, artist-created assets. But for bringing reality into the digital world—whether for VR, real estate, or film production—the era of the mesh is over. We are living in the age of the splat.

Sources

• Inria / Max Planck Institute: 3D Gaussian Splatting for Real-Time Radiance Field Rendering (SIGGRAPH 2023)
• Hugging Face Open-Source 3D Gaussian Splatting Implementations
• SIGGRAPH 2023 Technical Papers and Rendering Benchmarks