See https://github.com/KhronosGroup/glTF-Sample-Models/blob/main/2.0/FlightHelmet/README.md
I’ve been experimenting with Gaussian splatting to recreate the Flight Helmet model from the Khronos Group’s glTF sample collection.
The original glTF file weighed in at 43.06 MB. Using KTX2 texture compression, it dropped to 29.37 MB. My SOG-compressed Gaussian splat version comes in at just 3.6 MB.
My goal is to find the right balance between visual quality and file size. Smaller files mean faster load times, they get into the browser quicker, and ultimately, into the viewer’s hands and eyes faster.
For anyone who’s spent time on 3D optimisation, you’ll know it takes both time and expertise to get right. There are always trade-offs, compression artefacts, normal map inconsistencies, quirks in underlying meshes.
This is simply another way of tackling the same problem: how to deliver high-quality visuals that load fast and look great in the browser.
At the end of the day, especially for e-commerce, the average consumer doesn’t care whether they’re viewing a highly optimised glTF with advanced baking or a Gaussian splat. What matters is the visual experience.
The elephant in the room with Gaussian splatting, which is getting a lot of attention right now, is that albedo and lighting are baked into the object. There are extensions to the method that aim to address this, but compared to physically based rendering (PBR), you lose flexibility. You can’t just change materials or lighting and have a real-time renderer update the scene, at least not with the basic Gaussian splatting implementation.
www.neuralstorm.io