Search
On December 17, Moore Threads made a significant announcement regarding the complete open-sourcing of LiteGS, its proprietary 3D Gaussian Splatting (3DGS) foundation library. According to reports, when attaining image quality on par with the current top-tier solutions, LiteGS achieves an impressive training speedup of up to 10.8 times and slashes the parameter count by over 50%. When comparing models with an equal number of parameters, LiteGS's Peak Signal-to-Noise Ratio (PSNR) metric outperforms mainstream alternatives by 0.2 to 0.4 decibels (dB), while simultaneously cutting down training time by a factor of 3.8 to 7. For lightweight model implementations, LiteGS requires merely around 10% of the training time and 20% of the parameters of the original 3DGS approach to achieve comparable image quality.
-
C114 Communication Network -
Communication Home
