Larger file size of .splat files

[Bishadkoju]

Hi,
I was wondering why the file size of .splat files are so high (100+mb) compared to Kevin's .splat files(25mb).
Skimming through the codebase, this implementation also discards the spherical harmonics coefficient. Is there any specific reason for the larger file size ?
Thank you ! :)

[mkkellogg]

My splat files save the splat rotation components as 32-bit floats rather than 8-bit integers, so my splat files are indeed larger (I still plan on making that more efficient). However when you mention 25mb vs 100mb+, I think you might be comparing different files? Also keep in mind that for a lot of these scenes there are 2 different versions, one generated from a lower detail point cloud and one generated from a higher detail point cloud (iteration_7000 and iteration_30000 in the data set from the original paper respectively). As an example, my high detail version of garden.splat is 245mb whereas Kevin's high detail version is 186mb.

[hybridherbst]

Would be awesome if the optimized formats from Aras could work - the tooling already exists and allows a ~5-20x compression rate.
(Link: https://aras-p.info/blog/2023/09/27/Making-Gaussian-Splats-more-smaller/)

[mkkellogg]

@Bishadkoju If you're still interested, you can follow our discussions about a new universal .splat format here: #47

[Bishadkoju]

@mkkellogg yes i am following it. Thank you .