How important is the image resolution when creating gaussian splats? I will let you decide with this comparison video. Here is my methodology: The source is a 4k video drone shot of a beautiful church in Brazil. #Computervision #3D #VFX (1/3)

I extracted 350 frames at 4k, 1080p, and 540p. I processed camera poses for one set and then swapped the images for gaussian splat training. I trained all three scenes using Splatfacto-Big via Nerfstudio. Then, created an animation path which I used for all three scenes. (2/3)

How else could I show the comparison? Share comparison video still frames? This is part of the 100 days of splats. Today is day 3. Follow me on my journey as I post 100 splats in 100 days. (3/3)

That's a great comparision.

Is there a way to watch twitter videos without the low resolution and compression?

I think it depends on the device you view it on.

Worth knowing that there isn't much of a gain from to 4k, especially if you can get better low light capture or higher frame rate at HD

There was a definite fall off after 1080. The original Gaussian splat project by default capped images at 1600 pixels on the longest dimension. I feel like that’s around the inflection point where quality improvements fall off.

@3pointedit Yes, inria 3DGS capped the images to 1600 pixels. Since splatfacto does not capped to 1600 pixels, so the 4K in splatfacto is worth to try I think

@3pointedit You can uncap 3DGS. Just need to add a flag when running the commands.