Kyutai's AURA Talking Avatar Model: Real-Time Streaming Faces for Your Voice
Industry News & Updatespublished on 2026-11-22T00:24:35+08:00

Kyutai's AURA Talking Avatar Model: Real-Time Streaming Faces for Your Voice

AURA (AUdio-dRiven streaming Avatar) is Kyutai's family of speech-to-video models, released 2026-11-20 with the paper, code, weights and a 30,000-clip synthetic dataset. The claim that matters: a portrait plus a live speech stream becomes talking-head video in real time on a single GPU — the avatar follows the voice with low latency, chunk by chunk, instead of waiting seconds to denoise a finished clip. This page is an external reference to what AURA is and is not — last checked 2026-11-21.

At a glance

External referenceKyutai AURA — kyutai.org's release post, the GitHub repository (kyutai-labs/aura) and the Hugging Face collection
Availability on ImgoNot on Imgo, and not generatable from Imgo — the weight licenses forbid commercial use, so an Imgo integration would be off the table even technically.
Last checked2026-11-21
What you can generate on Imgo todayThe live Seedance 2 family renders talking-head and dialogue clips with native synchronized audio — scripted spots, UGC-style ads, presenter frames — quoted in credits before each run.
Official source and billing unitKyutai releases everything openly — there is no AURA API billing. Weights: AURA-1.3B under CC BY-NC 4.0, AURA-5B under Snap's non-commercial license.

Make talking-head video on Imgo today: Seedance 2.0 with native audio, in the Video Generator.

What AURA actually is

Two models, one idea. AURA 1.3B streams 576p talking-head video in real time — 20.7 frames per second for 16 Hz playback on one GPU, 23× faster than the teacher it was distilled from. AURA 5B streams 720p at 9.7 fps, matching its TalkVerse teacher's quality at 17× the speed. Both were distilled from 50-step bidirectional teachers into 4-step causal students that generate chunk by chunk with a rolling KV cache and no audio look-ahead: only the current chunk's audio is needed before its video starts rendering. That last property is the whole point — it is what makes an avatar that can hold a live conversation, not one that answers on a five-second delay.

Where the speed comes from

The release is a distillation study as much as a model: a two-stage pipeline that first regresses the causal student onto the frozen teacher's trajectory, then unrolls the student without gradients and matches teacher and critic scores on the student's own path. The engineering details — block-wise causal attention over chunks of latent frames, motion context carried entirely by the KV cache, per-chunk diffusion timesteps — are why the thing streams at all. The paper (BMVC 2026) carries the full method.

The licenses, stated plainly

Open does not mean unrestricted here. The code carries its own license — the weights are both non-commercial — AURA-1.3B under CC BY-NC 4.0, AURA-5B under Snap Inc.'s non-commercial license (its teacher descends from Snap's TalkVerse). For research, personal projects and prototypes, that is fine. For a product — a customer-support avatar, a monetized companion app — you would be building on weights you cannot ship. Read the license files on the repository before writing a line of code around them.

Who AURA is for, and who it is not

AURA is for people building interactive systems: a face for a voice assistant (the release demos it driving Kyutai's Unmute voice AI live), a streamed avatar for an agent, a research platform for latency work. It is not a production tool for finished marketing video — the output is a talking head at 576p/720p, tuned to follow audio fast, not to compose a three-shot ad with product beauty passes. If your deliverable is a scripted presenter, a UGC-style ad or a product spot with dialogue, a generation model with native audio is the right instrument, and latency stops mattering.

The Imgo path for talking-head content

On Imgo the live Seedance 2 family renders dialogue and presenter clips with native synchronized audio — the spoken line, the room and the effects are part of the same pass, with first- and last-frame control to hold a look across cuts and reference images to hold a face. Every run quotes its credit cost before you submit. You trade AURA's real-time conversation for finished-shot quality and a license that allows commercial work — which is the trade most marketing and content teams actually need.

FAQ

Is AURA open source?

The code, paper, weights and a 30K-clip dataset are published openly — but the weights are non-commercial (CC BY-NC 4.0 for 1.3B, Snap's non-commercial license for 5B). Open to study and prototype — not open to ship in a product.

Can AURA run live at conversational latency?

That is its design target: 576p at real time (20.7 fps) on a single GPU with no audio look-ahead. The public demo drives a live voice conversation.

Can I use AURA on Imgo?

No — it is not integrated, and its non-commercial licenses make an Imgo integration the wrong move. For talking-head video you can generate today, the live Seedance 2 family renders dialogue with native audio.

What is the difference between AURA 1.3B and AURA 5B?

Scale and resolution: 1.3B streams 576p in real time on one GPU — 5B streams 720p at 9.7 fps with quality matching its larger teacher. Same causal, few-step architecture — different speed/resolution trade-offs.

Sources: kyutai.org's AURA release post of 2026-11-20 (specs, distillation pipeline, demo), github.com/kyutai-labs/aura (code and license files), the Hugging Face collection (weight licenses, dataset) — all checked 2026-11-21.

Share

Related Posts

No image

Industry News & Updates

AURA is Kyutai's open-weights answer to real-time talking avatars: audio in, streaming talking-head video out, on one GPU — with non-commercial weight licenses that matter before you build on it. Checked 2026-11-21.

2026-11-22T00:24:35+08:00

No image

Industry News & Updates

Suno Visuals is Suno's song-to-video feature: clips up to 15 seconds, lyric-synced, credits sold separately. What it covers, what it can't, and the direct-your-own alternative on Imgo's live Seedance 2 family — checked 2026-11-21.

2026-11-22T00:24:19+08:00

No image

Industry News & Updates

LTX-2.5 is Lightricks' open-weights audio-video model: multi-shot continuity, synchronized sound, auto duration and a community license that allows commercial use. What that means in practice — checked 2026-11-21, with the live Seedance 2 family as the generate-today option on Imgo.

2026-11-22T00:24:05+08:00