You're welcome and thank you for trying it out! Hard at work to make performance better - the "main" kernels are at least as fast as gSplat, so now need to rem

You're welcome and thank you for trying it out! Hard at work to make performance better - the "main" kernels are at least as fast as gSplat, so now need to remove other overheads. That, and make splatting train more efficiently in general, lots of compute is wasted on small steps. Ps: the web version takes a minute to warm up and is generally slower, do try a native version if you haven't yet!