kyutai/AURA-1.3B: Real-time streaming talking-head video model

huggingface.co2 min read

Kyutai has released AURA-1.3B, a 1.3-billion-parameter streaming talking-head video model capable of generating real-time video from a portrait image and a speech audio stream. The model uses a rolling KV cache and a 4-step causal student distilled from a bidirectional teacher model conditioned on Mimi audio tokens. Released under a non-commercial license, the checkpoints and reference code are made available for research and non-commercial applications. Open article →

More on this topic