Custom NeRF render script [Now with Multi-GPU rendering!]

[JamesPerlman]

Hey there, I made a custom render_nerf.py script. It allows you to render a transforms.json with a camera path. It is fully compatible with the same style of json file that you can pass into run.py via --screenshot_transforms. It should be self-explanatory if you run python render_nerf.py --help

The main change that most people will be able to use is that it allows for multi-GPU rendering. It simply splits the render tasks into multiple subprocesses, each with the environment variable CUDA_VISIBLE_DEVICES set to a different value (up to your maximum number of physical devices). I'm certain there is a better way to leverage multiple GPUs, but this works for now.

It also adds the ability to animate the following properties, but these need to be added to the transforms.json manually or via this blender add-on that has no instructions yet: https://github.com/JamesPerlman/instant_ngp_tools

• aabb crop/bounds
• camera focal length

[itsmatthewc]

Hi, I'm excited to come across your pull request! I was looking to use this software, but I don't have a NVIDIA GPU. Does your pull request mean that I can use the software?

[JamesPerlman]

Hi, I'm excited to come across your pull request! I was looking to use this software, but I don't have a NVIDIA GPU. Does your pull request mean that I can use the software?

Unfortunately not, sorry! This is still only compatible with NVIDIA GPUs.

[scyheidekamp]

I want to change the aabb crop/bounds manually. Could you show me an example of what that would look like in the transform.json? Because I run into trouble trying to add your addon in Blender and exporting from there.

[riadas]

@JamesPerlman Not sure why this was closed, but this functionality would be super useful to me for an immediate use case! Should I still expect the script to work? I've trained some models using instant-ngp that are running at 5-10 FPS, and would love to take advantage of additional GPUs to get that closer to 60 FPS (probably a long shot, but it's worth a try!).