Procedural Gaussian splats in Houdini: six fits show gains from 12 to 24 training views

A developer rendered a small procedural valley scene in Houdini and trained Gaussian splats on the synthetic images produced. Using 4,096 splats and 600 training steps, raising the training views from 12 to 24 improved held-out quality for both seeds tested. Moving from 24 to 36 views added little full-frame PSNR and slightly lowered foreground PSNR. The findings apply only to one 192 by 192 scene and a fixed-capacity trainer, with fit times and memory figures reported for this tiny setup. The 36-view seed-381 result was chosen for delivery, and its compressed SPZ file was 76.7 percent smaller than the binary PLY while still loading in a browser viewer.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)
Log in to join the discussion and vote.
Log in