Neuralangelo Research Reconstructs 3D Scenes | NVIDIA vs DeepMotion

Neuralangelo is an NVIDIA Research model that reconstructs detailed 3D objects and scenes from ordinary 2D video. Feed clips filmed from multiple angles and the system builds textured meshes of buildings, sculptures, vehicles, and interiors suitable for games, digital twins, and robotics sims.

Earlier neural reconstruction methods blurred fine textures like roof shingles, glass, and marble. Neuralangelo extends instant neural graphics primitives from NVIDIA Instant NeRF to capture those surfaces with higher fidelity, so smartphone footage can produce assets creative teams can import into design tools without a laser scan or manual modeling pass.

Researchers, game artists, and robotics engineers use Neuralangelo to digitize real-world objects from Michelangelo-scale statues to campus buildings. The project is open source on GitHub under nvlabs/neuralangelo with no commercial license fees for the research code itself.

DeepMotion builds 3D character animation from text prompts or reference video through a browser, with no mocap suit required. Its two products, SayMotion and Animate 3D, turn written motion descriptions or uploaded MP4, MOV, and AVI clips into exportable skeleton data for game and film pipelines. Output ships as FBX, BVH, GLB, or MP4 files you can retarget to custom characters.

Traditional motion capture needs specialized hardware and studio space. DeepMotion runs markerless full-body tracking in the cloud, so indie studios and solo animators can produce fight choreography or dance moves from a phone video. SayMotion adds generative inpainting to extend or blend motions beyond stock libraries, which stock animation sites cannot match without manual editing.

Game developers, XR artists, film students, and educators use it to prototype character motion before committing to a full rigging pass. The Free Animation Credit Program lets users earn back credits by labeling motion descriptions or correcting poses in the Rotoscope editor.