TaoMate-HD: Progressive-Resolution Autoregressive Audio-Video Generation

In plain English

Alibaba's TaoLive team generates a streaming digital human at high resolution by denoising each short chunk twice at low resolution to fix the coarse structure, then once more to add detail, instead of denoising at full resolution throughout. On the same model and hardware this runs 1.6x faster than direct high-resolution generation, and a stage-parallel deployment adds a further 1.9x. Each stage also passes its own audio-visual prediction forward as history, which cuts character error rate by 11.18 percentage points and stops speech repeating or cutting off when the script changes mid-stream — the failure mode that breaks live avatar demos.