[Hackathon] NeRF in Mojo

Community Showcase
1

TL;DR: Iimplemented a Neural Radiance Fields (NeRF) rendering pipeline in both PyTorch and Mojo, including ray sampling, neural network inference, and volume rendering. PyTorch is currently 73.5x faster, but the project taught me a lot about Mojo and its potential.

What I Built

Complete End-to-End NeRF Implementation: Not just the neural network, but the entire 3D rendering pipeline:

  • Ray Generation: Camera rays for each pixel

  • Point Sampling: 3D points along rays

  • NeRF MLP: Neural network for density + color prediction

  • Volume Rendering: Alpha compositing integration

  • Image Synthesis: Final pixel color computation

Both frameworks render the same algorithm identically - this isn’t just a toy example, it’s a real implementation that generates real results (24.59 dB PSNR on the class Lego scene).

Try It Yourself

Clone the github repo and test it for yourself

  • Memory Optimization: Novel approaches to graphics pipelines
3 Likes
2

This topic is very hot in the GIS space right now.

Your title says " PyTorch is currently 73.5x faster, but the project taught me a lot about Mojo and its potential." but the readme.md says mojo+gpu is 10,077x faster than PyTorch GPU NeRF. Bit confused. Are those two separate benchmarks that aren’t comparable?

4

the link is not working.. can you pl share updated link. super interested in how LumaAI’s nerf were implemented and to compare with mojo implementations