As a footnote in quite a lot of the Gaussian Splatting reviews, we see the disclaimer that the resulting file sizes are quite large. For most of the people that are looking into 3DGS right now, that’s not really a deal breaker for them, but for any hope of it catching onto consumers at large, that will need to change.
Aras Pranckevičius has been working on a few exciting Gaussian Splatting explorations recently and published an article yesterday about reducing file size.
You might recognize Aras Pranckevičius’s name from last week, when he was able to get 3D Gaussian Splatting running in Unity. While I could summarize everything that Aras says, I think he does a great job explaining it in blog post here. In short, Aras was able to drastically lower the file size of a Gaussian Splat.
Honestly, I can’t really tell the difference between the two unless I really begin to zoom in. Aras demonstrates several different iterations and file sizes, with associated quality attached to them. If you don’t care about the visual fidelity, he can reduce the file size by as much as 19 times! However, as he is the first one to admit, it doesn’t exactly look great.
The gaussian splatting data size (both on-disk and in-memory) can be fairly easily cut down 5x-12x, at fairly acceptable rendering quality level. Say, for that “garden” scene 1.35GB data file is “eek, sounds a bit excessive”, but at 110-260MB it’s becoming more interestingAras PranckevičiusTweet
There has been some exploration with optimizing NeRF file sizes, with Binary Radiance Fields greatly shrinking them. NeRFs still hover around 35mb (INGP) to 55mb (Luma), it still is a welcome decrease from the standard gigabyte file size we have seen thus far from 3DGS.
Check out Aras’s original article for a larger deep dive into the mechanics behind it!
