Slop TVNewsLatest
News

Kyutai AURA streams talking-head video in real time on one GPU

The open release pairs a 1.3B model that streams 576p at 20.7 FPS with a 5B model at 720p, alongside code, weights and 30,000 clips.

Illustration: Kyutai AURA streams talking-head video in real time on one GPU
Illustration: SLOP TV News. Logos are trademarks of their owners.

Key takeaways

  • Kyutai released AURA, a family of causal speech-to-video models that stream talking-head video chunk by chunk, in a blog post dated November 20, 2026.
  • AURA 1.3B streams 576p video at 20.7 FPS in real time on a single GPU, 23 times faster than its bidirectional teacher, while AURA 5B streams 720p at 9.7 FPS, 17 times faster than its TalkVerse teacher.
  • Kyutai released the paper, the code, the weights of both models and a dataset of 30,000 synthetic videos with precomputed teacher trajectories, plus a live demo at aura.kyutai.org.
  • In Kyutai's own human study, both AURA models were preferred over Hallo3 on all four axes, even though Hallo3 led on two distribution metrics.

Kyutai released AURA, a family of speech-to-video models that generate talking-head video chunk by chunk and stream it in real time, in a blog post dated November 20, 2026. The work is a collaboration with Sorbonne Université and Google DeepMind and is being presented at BMVC 2026.

AURA stands for Audio-driven streaming Avatar. The problem it attacks is latency. Audio-driven video diffusion models usually need dozens of denoising steps over multi-second clips, and the audio of the whole clip before they can start, which rules them out of a live voice assistant. AURA instead distills 50-step bidirectional teachers into 4-step causal students that run with a rolling KV cache and no audio look-ahead. Only the current chunk's audio is needed to start generating.

Two sizes ship. AURA 1.3B streams 576p video in real time on a single GPU at 20.7 FPS, faster than its own 16 Hz playback rate and 23 times faster than its bidirectional teacher. AURA 5B streams 720p at 9.7 FPS, a 17 times speedup over its TalkVerse teacher, or 30 times on the diffusion transformer alone. Both were measured on a single H100 GPU.

The 1.3B model is distilled from a 1.3B teacher trained with the TalkVerse pipeline on Wan2.1 1.3B, with audio features from Mimi, Kyutai's streaming audio codec, instead of Wav2Vec2. AURA 5B is distilled from TalkVerse (5B). AURA 1.3B is also the face behind Unmute, Kyutai's voice AI: pick a face or upload your own, click to talk, and the avatar is generated live from Unmute's speech. The demo model was retrained with smaller 2-frame chunks of 0.5 seconds to cut the avatar's lag behind the voice to about 1.2 seconds, from about 2 seconds with the paper's 4-frame chunks, while staying real time.

The release is open science. Kyutai published the paper, the code at github.com/kyutai-labs/aura, the weights of both models, and a dataset of 30,000 synthetic videos with their precomputed teacher trajectories on Hugging Face, plus a live demo at aura.kyutai.org.

One honest note from the paper. These speeds are Kyutai's own measurements. Hallo3 topped two distribution metrics, FID and JEDi, on both test sets, but the authors write that qualitative inspection shows exaggerated mouth motion and recurring hand artifacts, and in their own human study of 400 pairwise comparisons by 27 raters, both AURA models were preferred over Hallo3 on all four axes.

Try AURA at aura.kyutai.org, or pull the code and weights from the Kyutai GitHub repository and its Hugging Face collection.

Sources

  1. kyutai.org - primary blog post: all specs, the release and the demo
  2. arxiv.org - the paper abstract and its November 19 submission date