Store all 3DGS in RAM and then move the required 3DGS to video memory

[smart4654154]

in the parper, I find:
We update the cut every 2 frames to add details by transferring nodes to the GPU from CPU RAM, and run cleanup every 100 frames.
You put all the 3DGS in RAM, and when rendering, call the required 3DGS, then put it in the VRAM, and perform rendering.
Do I understand correctly?

Thank you very much.

[venite-xjc]

Hello! My understanding aligns with yours, but I didn't grasp how this paper efficiently schedules data while achieving real-time rendering. The code only demonstrates a method of loading all data into GPU memory at once. Do you have any thoughts on the implementation?

[laphisboy]

i am also very curious about this. could you point out where in the code the CPU to GPU Gaussian transferring happens, if it is implemented?