MonoFlex3D: Low-Latency Glasses-Free 3D Teleconference from a Single Camera

The transformative potential of immersive 3D teleconference to replicate face-to-face interaction remains largely unrealized due to persistent barriers in usability and real-time performance. Existing systems often compromise on ease-of-use, require complex multi-sensor setups, struggle with the latency-fidelity trade-off in dynamic human reconstruction, or necessitate cumbersome wearable devices. This paper introduces MonoFlex3D, a novel real-time system designed to dismantle these barriers, enabling low-latency, glasses-free 3D teleconference using only a single commodity RGB camera. At its core, MonoFlex3D employs an avatar-driven approach with a compositional 3D Gaussian Splatting representation. This technique models the human body and face separately using parametric templates, with targeted refinement of the crucial facial region to capture fine-grained expressive details. Extensive evaluations of the MonoFlex3D prototype demonstrate superior real-time capabilities: it achieves high-fidelity avatar animation with approximately 10ms driving latency while maintaining an end-to-end system latency below 100ms at 30 frames per second. It also reduces 50.2% end-to-end latency versus the simplest capture systems, and enhances visual quality by 9% to 14.5% over state-of-the-art alternatives, all while drastically simplifying capture requirements. MonoFlex3D thus offers a significant step towards truly accessible and practical immersive communication.