Adventures in Gaussian Splatting: The Start of a New 3D Rendering Empire?

So, here’s the situation. I was happily plugging away at a work project involving point clouds in Unreal Engine, but very quickly, the universe reminded me that life is never that easy. Imagine you’ve just been handed a giant bag of tangled Christmas lights, but instead of lights, it’s a dataset the size of a small moon. Sure, point clouds sound great, but after attempting to load this behemoth, I started to wonder if the poor machine was reconsidering its life choices.

Point clouds are like that friend who insists they’re low-maintenance but still need you to hold their hand through everything. They eat up storage like it’s a buffet, and don’t even get me started on processing power. My work colleagues old laptop gave up faster than I do when someone mentions “hot yoga.”

Naturally, the alternative was converting this giant, noisy cloud into a mesh—because who doesn’t love turning random points into a giant mess of connected polygons? Spoiler alert: it wasn’t great. Imagine trying to build a house of cards in the middle of a windstorm—some points were missing, others were just outright wrong, and there I was, trying to coax them into something vaguely resembling geometry. Honestly, even after all that work, the results are still worse than my own homemade photogrammetry project from two summers ago ((24) Adventures in Photogrammetry. | LinkedIn).

But there could be a hero to this story: Gaussian Splatting. It’s like point-based rendering, but with a little more “oomph.” Think of it as point clouds with a glow called "splats". Each splat spreads its inspiration like a YouTube influencer, smoothly blending with its neighbors. The result? A system that delivers stunning renders in realtime.

But does it actually work? There’s only one way to find out: science! And by science, I mean throwing it at the software and crossing my fingers. Enter Jawset’s Postshot—the plucky beta software that allows you to feed it video footage, and voilà, Gaussian splats everywhere. The best part? It’s currently free (because it’s in beta, and, let’s be honest, free stuff always feels magical).

I didn’t make this easy. I wanted to challenge it, so I picked a park/woodland scene—the kind of environment that photogrammetry looks at and immediately screams, “Nope.” Dense foliage, reflective water, the kind of detail that makes 3D software weep. If this didn’t break Postshot, nothing would.

I fed it a lengthy 10-minute video of the above brook (because why not overkill?), and then, like any good scientist, I left it to crunch numbers while I went to grab a coffee. After some processing time (okay, a lot of processing time), I was finally greeted with the splat map results. To say that I was impressed would be an understatement. I’m talking jaw-on-the-floor levels of good. No more messy meshes, no UV mapping, and no need to sacrifice a small animal to the 3D gods just to get something workable.

The Benefits? Oh, Let Me Tell You.

For one, Gaussian splatting is faster and takes up less memory. It’s like switching from a dial-up to broadband—you’re never going back (Okay so I'm exaggerating here). And unlike point clouds that just kind of sit there, pretending to be real objects, splats let light dance around them in real-time. I swear, I watched shadows and light play with the brook’s surface like it was The Lion King on Broadway.

The smoothness of the rendering was absurdly satisfying. The foliage and water? They looked amazing. Photogrammetry would’ve thrown its hands in the air and walked away from all that running water. But not Gaussian splatting—it handled the chaos like a pro.

But It’s Not All Rainbows and Unicorns

It wasn’t perfect, though. Like its point-cloud cousins, Gaussian splatting comes with its own baggage. It had a few "floaters" (random points that show up uninvited and decide they’d rather do their own thing, thank you very much). And the edges of the reconstruction? Yeah, that’s where things got a bit dicey, like an unfinished puzzle where the pieces start making stuff up. It was the software’s best guess at filling in missing data, but let’s just say it’s more of an abstract art piece than a rendering masterpiece.

Oh, and don’t even think about trying to simulate reflections, refractions, or translucency—those splats don’t play well with traditional material properties. It’s like trying to make a cat do tricks—good luck with that.

So, Is Gaussian Splatting the Future?

While traditional mesh-based rendering isn’t going anywhere (because let’s face it, change is hard), Gaussian splatting is carving out a nice niche for itself. It offers a hybrid approach that could streamline workflows and make those big, memory-hungry datasets a little less terrifying.

And with developers continuing to push this technology forward, we might just see Gaussian splatting take over as a go-to technique for high-quality real-time rendering. Until then, it’s a tool worth adding to the 3D toolbox. I mean just check out the 3D splats below.

So, what do you think? Ready to splat your way into the future of 3D rendering?