Forward pass of rectified flow model

[vevenom]

I'm reimplementing the forward process of your rectified flow model but I am not able to find details in the paper and the implementation.

Could you elaborate more regarding converting a mesh into the active voxel representation and how you generate the noise for the rectified flow model during sparse structure generation and later during structured latents generation? Also, could you provide hyper-parameters that you use during the process?