search

r4dl / StopThePop

Original reference implementation of "StopThePop: Sorted Gaussian Splatting for View-Consistent Real-time Rendering"
https://r4dl.github.io/StopThePop/
Other
141 stars 6 forks source link

CUDA error: an illegal memory access was encountered #2

Closed ronly2460 closed 3 months ago

ronly2460 commented 3 months ago

Hi, when I try images at 3k, it will be run out of memory.

Is there anyway to fix and reduce memory size of your rasterizer?

This did not happen in the original rasterizer of 3D GS

steimich96 commented 3 months ago

Compared to 3DGS, our method should not consume that much more memory, so I guess you were pretty on the edge to begin with? And you use a GPU with 24 GB VRAM?

One option is of course to downscale your images before training. You can call convert.py with --resize to achieve this. For other tips, see the 3DGS FAQ