AURA: AUdio-dRiven streaming Avatar
Paper • 2611.26540 • Published • 2
kyutai.org/aura - AURA is a 1.3B speech-to-video generative model that runs in real time on a single H100 GPU at 576px resolution.
Note AURA 1.3B built on top of WAN2.1 1.3B with Mimi audio latents
Note AURA 5B built on top of WAN2.2 5B with Wav2Vec audio latents
Note Generated videos and predicted teacher ODEs used to train the Stage 1 of our 1.3B and 5B AURA models
Note Bi-directional 1.3B teacher for AURA, trained on top of WAN 2.1 T2V via LoRA using the training pipeline as in [TalkVerse](https://arxiv.org/abs/2512.14938)