NVIDIA’s neural surface reconstructor eats VRAM for breakfast

What it does\nNeuralangelo takes a sequence of video frames—assuming known camera poses—and reconstructs a detailed 3D surface using neural fields. It then extracts an explicit isosurface mesh, optionally textured, suitable for standard graphics pipelines. The codebase is built on NVIDIA’s Imaginaire library and supports multi-GPU training and extraction.\n\nThe interesting bit\nMany neural rendering projects stop at synthesizing views; this one recovers real geometry you can export as a mesh. The trade-off is a fierce appetite for hardware: the default configuration demands at least 24 GB of GPU memory, and the authors note that tuning it down to fit smaller cards sacrifices reconstruction quality.\n\nKey highlights\n- Official CVPR 2023 implementation from NVIDIA.\n- Exports textured or untextured isosurface meshes rather than just implicit fields.\n- Built on the Imaginaire library; supports multi-GPU training and mesh extraction.\n- Includes a Colab notebook for trying without local setup.\n- Requires known camera poses; preprocessing uses COLMAP and follows the Instant NGP JSON format.\n\nCaveats\n- The default configuration requires at least 24 GB of VRAM; running on 8–16 GB cards means dialing back quality.\n- It assumes accurate camera poses up front, so if COLMAP drifts, your mesh drifts too.\n- Business use requires filling out NVIDIA’s research licensing form.\n\nVerdict\nWorth a look if you are a researcher or graphics developer who needs high-quality mesh extraction from video and owns a data-center GPU. Everyone else should probably admire the teaser GIF and wait for a more efficient port.