Live Avatar: Streaming Real-time Audio-Driven Avatar Generation with Infinite Length

In plain English

This tackles the hardest part of real-time avatar video: keeping a large diffusion model streaming indefinitely without the face drifting or degrading. The authors distil a 14-billion-parameter bidirectional model into a causal few-step one and spread the denoising steps across GPUs as an asynchronous pipeline, reporting 45 FPS and 1.21 seconds to first frame on five H800 GPUs, with stable generation past 10,000 seconds; they also release a long-form benchmark, GenBench. The figure worth holding onto when costing a deployment is the hardware: this quality of infinite-length streaming currently assumes a five-GPU server behind each stream, and all numbers are the authors' own.