Working Log : optimizing Gsplat

April 11, 2025

A working log documenting experiments to optimize the training process for Gaussian Splatting (gsplat). Currently, using my COLMAP pipeline on a 5-minute 1080p video completes training in 30 minutes on an H100 GPU. I'll keep updating this.

April 19 , 2025

Switched to VGG-SfM instead of COLMAPS , shaving off 2 mins . Unlike COLMAP’s traditional pipeline with discrete steps, VGG-SfM trains nn components to perform each stage of SfM and integrates them into a single differentiable framework.