Normal estimation using gradient of density

[mra-h]

Hello,

Thank you so much for this implementation of ref nerf!

I have a question regarding on the gradient of the density to compute the normals. Why is raw_density.sum() used as outputs in torch.autograd as shown here.

Is it different than calculating the gradient as torch.autograd.grad(raw_density, means, grad_outputs=torch.ones_like(raw_density), retain_graph=True). ?