Meta and NVIDIA Optimize Muse Realtime Avatar for Efficient Video Streaming

Meta unveiled Muse Realtime Avatar, a real-time embodiment system that generates expressive video avatars synchronized to speech, with NVIDIA-optimized model efficiency enabling 12 concurrent sessions per GPU. The collaboration between Meta's AI Research team and NVIDIA focused on reducing computational overhead in the model's forward pass, deploying serving optimizations including persistent key-value caches, dynamic batching, four-bit quantization-aware training, and fused CUDA kernels.