NVIDIA turns smartphone video into detailed 3D scenes

Neuralangelo is an open-source 3D reconstruction model from NVIDIA Research that turns standard 2D video clips into high-fidelity 3D assets.

All you need is a multi-angle video shot on a regular smartphone. The model translates that footage into lifelike virtual replicas of physical objects, sculptures, and full-scale buildings.

What it can rebuild

Previous AI models that handle 3D reconstruction usually choke on tricky surfaces like repetitive patterns, homogeneous solid colors, or dramatic lighting shifts.

Neuralangelo largely fixes these blind spots. It reconstructs complex surface textures that used to break older photogrammetry tools, including roof shingles, glass panes, and polished marble.

In its public demos, NVIDIA showed the tool recreating Michelangelo’s David statue with delicate anatomical precision, alongside everyday objects like a flatbed truck and an entire park at NVIDIA's Bay Area campus.

Once generated, creative teams can export these 3D objects straight into standard design applications for game development, digital twins, and robotics simulations.

How it works

Under the hood, Neuralangelo relies on instant neural graphics primitives, the same foundational technology behind NVIDIA Instant NeRF.

The pipeline mimics an artist sculpting from stone. First, the AI selects multiple frames from the video to figure out the exact camera position of each viewpoint.

Next, it carves out a rough 3D representation of the scene to lock in scale and volume.

Finally, the model sharpens the render through iterative optimization, progressively refining fine surface textures until they match the original video.

The project was presented at CVPR 2023, and the official code is already available on GitHub.

Why it matters

We think this is a huge practical leap for anyone building virtual worlds. Instead of buying expensive photogrammetry rigs or hand-crafting 3D geometry from scratch, independent developers can now capture game-ready environments with just a smartphone.

While earlier NeRF experiments produced pretty visual viewpoints, Neuralangelo actually yields structured, high-fidelity geometry with usable material textures. It cuts the friction between the real world and digital engines down to almost zero.