search

google / mipnerf

Apache License 2.0
894 stars 109 forks source link

Why the training time is longer than NeRF #34

Closed Harper714 closed 2 years ago

Harper714 commented 2 years ago

Hi, It seems that mip-NeRF takes much longer time to train compared to NeRF. Is there any explanation or comparison? Thanks!

jonbarron commented 2 years ago

Hi, I'll need more information to advise. With the same model size and batch size, and running on the same accelerator, mip-NeRF should be comparable to or slightly faster than a vanilla NeRF implementation. If it's much longer you probably have a bug somewhere.

Harper714 commented 2 years ago

Thanks for your reply. I tested again carefully and found you are right. My mistake was caused by a different setting.

BTW, It seems that mip-NeRF need more GPU memory during training. I want to ask why and if it is possible to address this problem?

jonbarron commented 2 years ago

mip-NeRF does more math than NeRF when ray-casting (points along the ray have sizes / covariance matrices associated with them) so it should take more memory.

Harper714 commented 2 years ago

Many thanks for answering.