NeRF, rewritten in PyTorch, runs 1.3× faster

What it does

NeRF represents a 3D scene as a tiny fully connected neural network—roughly 5 MB of weights—that maps a 5D coordinate (position plus viewing angle) to color and opacity. This repo is a line-by-line PyTorch port of the original TensorFlow code; it trains the same network architecture so you can synthesize novel views of complex scenes from a handful of input photos. The README says it has been numerically tested against the official implementation and comes with pre-trained checkpoints for standard benchmark scenes like Lego and Fern.

The interesting bit

The selling point here isn’t novelty—it’s fidelity with a slight performance edge. The author claims a 1.3× speed-up over the original TensorFlow version while keeping the training loop and outputs numerically identical, which is harder than it sounds when moving between frameworks. There is even a separate reproduce branch containing tests that verify every function matches the reference implementation.

Key highlights

• Claims 1.3× faster training than the official TensorFlow release
• Numerically tested against the original authors’ code
• Includes pre-trained models and configs for synthetic (Lego, Ship) and real-world LLFF (Fern, Flower) datasets
• Reproducibility branch with py.test suites to verify parity
• Network weights are tiny (~5 MB) for a full scene representation

Caveats

• The README pins PyTorch 1.4, which is several years old; compatibility with modern torch versions is unclear
• Training your own real-world scenes requires external dependencies like COLMAP and the LLFF codebase, not just this repo
• “Low-res” training still takes ~4–8 hours on a single RTX 2080 Ti, so it is hardly lightweight by modern standards

Verdict

Grab this if you need a battle-tested NeRF baseline in PyTorch for research or comparison. Skip it if you are looking for the latest accelerated NeRF variants—this is the 2020 original, faithfully preserved.