3DTopia / LGM

[ECCV 2024 Oral] LGM: Large Multi-View Gaussian Model for High-Resolution 3D Content Creation.
https://me.kiui.moe/lgm/
MIT License
1.71k stars 118 forks source link

CUDA Memory when rasterization #80

Closed wang-qinliang closed 2 weeks ago

wang-qinliang commented 2 weeks ago

Thank you for the excellent work of the author. Let me describe the problem. When the number of Gaussian points is 65536, the rendering operation takes up approximately 2.25GB of video memory. When the number of Gaussian points is 131072, the rendering memory is 2.4GB. However, when the number of Gaussian points is 196608, the video memory suddenly increases to 4.28GB. Theoretical analysis shows that it will not cause a significant amount of video memory occupation, and the renderer does not have any problems. Have you encountered this problem?