At NeurIPS, NVIDIA announced a wave of open physicalAI tools, but one stands out for anyone working with neurally reconstructed 3D scenes and radiance field based pipelines: Omniverse NuRec Fixer, a new model designed to rapidly correct artifacts common in NeRF and Gaussian Splatting workflows based upon Difix3D+.
NuRec Fixer targets issues such as blurs, holes, and missing geometry in novel viewpoints, often caused by sparse capture, motion blur, or sensor noise. Built on NVIDIA Cosmos Predict, the model can repair these artifacts nearly instantaneously, enabling more stable and production-ready datasets for simulation, robotics, and digital twin environments.
For teams working with Gaussian Splatting, where imperfect input data often produces distracting gaps or noisy regions, NuRec Fixer has the potential to dramatically improve usability, especially when preparing scenes for simulation in Omniverse, AV testing, or robotics training. Excitingly, this version also comes with a commercially permissible Apache 2.0 license.
The tool is part of NVIDIA’s expanding Cosmos physical-AI ecosystem, which also includes:
LidarGen, a world model that generates synthetic lidar for AV simulation.
Cosmos Policy, for converting large video models into robust robot policies.
ProtoMotions3, a GPU-accelerated framework for training realistic digital humans and humanoid robots.
The company also highlighted more than 70 NeurIPS research contributions and new benchmarks recognizing the openness of the Nemotron model family.
While autonomous-driving model Alpamayo-R1 and new speech-AI tools received significant attention, NuRec Fixer is a practical breakthrough for radiance field creators and companies, promising cleaner splats, more reliable simulations, and faster iteration across Omniverse based pipelines.
NVIDIA has created a Hugging Face space for Fixer, in addition to also publishing the code to GitHub.
