Why is NeRF very slow at test time and how to make it faster ?
Jiakui opened this issue · comments
I heard that NeRF is very slow at test time, "Rendering a single 1280x720 image on one such GPU takes ∼30 seconds at test time ".
Can we just generate low resolution images and then use some super resolution method to improve the resolution ? That may make it faster.
Hello @Jiakui , in my opinion, the biggest benefit of NeRF's rendering process is the guarantee of multi-view consistency. Applying a super-resolution algorithm on top of each image separately will lose that potentially.
OK, can you briefly explain why NeRF runs very slowly at inferece stage ? Is it due to the MLP layers computation cost too high, or the volume rendering stage very slow ? Thanks!
These two are all reasons, but I think the volume rendering part is more problematic. For a 400x400 image, it samples 64 coarse samples and 64+128 fine samples, which makes 400x400x(64+64+128) = 40960000 samples to do MLP inference.
To accelerate this, you can for example try NSVF that uses a special data structure to accelerate.