Conversation

OH MY GOODNESS. I can now render multiple NeRFs in the same scene with my port of

. This was a long haul to rewrite the renderer and it’s not yet perfect but LOOK AT THEM! All overlappy and composited together.

0:40

36K views

193

1,492

Replying to

and

What are your thoughts on the nvidia license?

Replying to

and

Right now I’m just using the standard ngp license for noncommercial research & evaluation. It’s more prohibitive than I’d like but NVIDIA is a huge company and they gotta protect their IP.

Replying to

have you looked into the nerfacc github.com/KAIR-BAIR/nerf library? it implements some of the same techniques in NGP, its what nerfstudio uses and is MIT licensed

github.com

GitHub - KAIR-BAIR/nerfacc: A General NeRF Acceleration Toolbox in PyTorch.

A General NeRF Acceleration Toolbox in PyTorch. Contribute to KAIR-BAIR/nerfacc development by creating an account on GitHub.

7:40 PM · Nov 21, 2022

Replying to

I might end up going this route, but for now I need maximum control and performance provided by CUDA. Some features I plan to add (multi-GPU support, volumetric distortion modifiers, mask shapes) would be slowed down by PyTorch’s abstraction layers, or am I wrong?

Replying to

ah yeah, nerfacc does have CUDA implementations of its renderer components, so its still about as fast as NGP, but its definitely built to be used through pytorch and might not be easy to separate from it

Show replies

Replying to

Also one of the researchers behind

@LumaLabsAI

created alexyu.net/plenoctrees/. Im curious about doing everything inside a shader.