Over the weekend, NVIDIA quietly released the code for one of their most exciting papers from last year: QUantized Efficient ENcoding of Dynamic Gaussians for Streaming Free-viewpoint Videos—or simply, QUEEN.
QUEEN reconstructs dynamic captures from multi view cameras into lifelike 3D using Gaussians. The framework blends fast training, efficient memory usage, and high fidelity rendering to bring dynamic scenes to life, with playback speeds reaching up to 350 FPS.
For those who attended this year’s GTC, you might recognize QUEEN as the method powering their incredible 3DGS VR experience in the XR Pavilion. The method comes out of NVIDIA’s AMRI Lab, led by Dr. Shalini Di Mello.
QUEEN is released under NVIDIA’s proprietary license, so you’ll need to contact their team for any commercial use. It also relies on the original Inria 3DGS implementation, which requires a separate license. But for research or personal experimentation, QUEEN is an excellent option for compact, real time representations of dynamic radiance fields.
Excitement around dynamic 3D seems to be reaching a new high this past week. It feels like radiance field representations are finally tipping into the mainstream, where the use cases are as inspiring as the tech itself.
You can explore the QUEEN project page for video demos or access the code through the NV Labs GitHub. For those heading to CVPR this week, you can see AMRI Lab director Dr. Di Mello speak at the Workshop on Photorealistic 3D Head Avatars.
