3D view synthesis, now with 100% fewer neural nets

What it does Plenoxels reconstructs a 3D radiance field from ordinary photos so you can render unseen viewpoints. It skips the neural-network approach and instead performs voxel optimization directly, producing a scene representation you can render from new angles.

The interesting bit The project replaces a network with a data structure—a voxel grid—though the repo name svox2 is deliberately not positioned as a successor to the earlier svox work. The README is admirably honest about what is missing: there is no GUI viewer, training still uses FP32, and the authors suggest ReLU activations may be giving surfaces a semi-transparent look. It is research code that knows it is research code.

Key highlights

• Supports multiple standard dataset formats (NeRF-Blender, LLFF, NSVF, CO3D) and auto-detects which one you threw at it.
• Includes a parallel task executor (autotune.py) for scheduling batch training, ablations, or hyperparameter sweeps across multiple GPUs.
• Handles custom 360° captures via COLMAP, though the README warns that radial distortion can blur results and recommends --noradial.
• Ships with a CUDA extension; the authors note that using ninja makes compilation tolerable.

Caveats

• Windows is not officially supported; the authors have only tested on Linux.
• No support for the instant-ngp dataset format (the project predates it), so you will need to run the provided conversion script.
• Several quality improvements remain unimplemented, including FP16 training, exposure modelling, and a mip-nerf 360 distortion loss to reduce floaters.

Verdict Grab this if you are researching novel-view synthesis and want a voxel-based baseline to hack on. Skip it if you need a polished, end-user application—the missing GUI and manual COLMAP workflow make it strictly a research tool.