Questions about training t * 1000 scaling in Rectified Flow and other details

Hi,

First of all, thank you for this amazing work!

I’m reading your code on Rectified Flow and had a small question.
In the training code, I noticed that the time step t is multiplied by 1000 before being fed into the model (t * 1000). I’m curious about the reasoning behind this scaling.

Could you please share any references or prior works you followed when using this scaling? I’d love to understand it better.

And when I run datatoolkits/voxelize.py, I encounter the following issues:

Segmentation fault occurs in:
o3d.create_from_triangle_mesh_within_bounds

AttributeError occurs in:
utils3d.numpy.io.write_ply

Error: module 'utils3d.numpy' has no attribute 'io'

Are there alternative methods or updated functions to replace these calls?

Thanks so much for your time and help!

Best regards,

Start with the training code that passes t * 1000 to the model, then inspect datatoolkits/voxelize.py and the reported calls to o3d.create_from_triangle_mesh_within_bounds and utils3d.numpy.io.write_ply. Determine the intended scaling rationale and identify compatible replacements or fixes for the segmentation fault and missing utils3d.numpy.io attribute; done means both questions have documented answers and voxelize.py runs without these errors.