Open zjcs opened 1 month ago
I've tried that before: https://github.com/wanmeihuali/taichi_3d_gaussian_splatting/pull/53 It seems taichi's autodiff system is still buggy and cannot produce the correct gradient. It seems the taichi's autodiff implementation is pretty naive.
When I read the taichi doc, taichi support autodiff funcation by kernel_func.grad. So is it possible to directly use kernel_func.grad in the backward function during taichi-3dgs?