Atmee V2 vs Seedance: How Causeur and a Real-Time Avatar Model Render in Seconds

Atmee V2 is a real-time video model for AI avatars, launched on 4 November 2026 by Atmanity, the Santa Clara lab behind atmee.ai. It turns one photo and about ten seconds of voice into a character you can video-call: it talks, listens, gestures and reacts while you speak. It can also render ordinary short clips. On the launch demo, the same 10-second clip took Seedance 2.5 233.6 seconds and $0.50, and took Atmee V2 3.5 seconds and under one cent.

I went through the launch video, Atmanity's research posts, the API docs, the public benchmarks and the developer tooling. This post covers what Atmee V2 is, how its two models (Causeur and the V2 renderer) fit together, how it compares with Seedance, what the API looks like underneath, and which claims are still unverified. Every clip below is hosted on this site, cut from Atmanity's own launch material, so it keeps playing if a social post disappears.

Atmee V2 at a glance

• What it is: a real-time, full-frame video foundation model for live AI avatars, plus a behavior model called Causeur.
• Input: one image and 10 to 15 seconds of voice.
• Output: a live video call, a voice-only call, or an mp4 clip of 1 to 15 seconds.
• Speed claim: up to 50x faster and 10 to 100x cheaper than Seedance 2.5 and Wan 3.0, running live on one GPU.
• Demo numbers: 3.5 s vs 233.6 s for a 10 s clip; 6.4 s and $0.01 vs 331 s and $1.19 for an 8 s clip.
• Price: 2 credits a minute for a full V2 conversation, about $0.21 to $0.30.
• Access: invite code or the Growth plan ($49 a month) and up.

What is Atmee V2?

Atmee V2 is the second avatar engine behind atmee.ai. Atmanity calls it "a real-time, full-frame video foundation model for live avatars". Full frame is the important word: V1 animated a talking head from the shoulders up, while V2 generates the whole shot, face, body, hands and surroundings, in square, 16:9, 9:16, 4:3 and 3:4.

The input is one image plus a short voice sample (the founder says 10 seconds, atmanity.io says 15, and the API accepts 10 to 60 seconds for a voice clone). The output is one of three things: a live video call, a voice-only call, or a generated mp4 clip.

Watch the clip with sound

A live Atmee V2 call on a tablet, from Atmanity's 4 November 2026 launch video. The avatar answers with expression and hand movement, not only lip sync.

How does Atmee V2 work? Causeur decides, Atmee V2 renders

Atmee V2 is two models, not one. In the launch video the founder puts it simply: "we built two models, a body and a brain". Atmanity's site sums it up as "Causeur decides, Atmee v2 renders."

The two-model stack. Causeur is the brain that decides what happens; Atmee V2 is the body that draws it.

What is Causeur?

Causeur is Atmanity's behavior model, "the behavior model that brings an agent alive". It handles personality, emotion, timing and active listening, moment to moment. In the launch video it is "a model that brings context and behavior into the voice, the body, and the response". Concretely, Causeur decides:

• When to speak. Whether the user has finished, paused to think, or is interrupting, and whether to answer, wait, or give a short reaction instead of a full reply.
• How to feel. Which emotion and expression fit the moment, so the renderer smiles at the joke and not three seconds before it.
• What to do while listening. Nods, small smiles and "go on" signals while the user talks, which is what makes a call feel like someone is there.
• When to act on the app. In a session with app tools, Causeur is the part that calls them (a v1 avatar accepts tools but never calls them).

Atmanity benchmarks Causeur against full-duplex speech models on three turn-taking cases, higher is better:

• User ends a turn: Causeur 0.92, Moshi (Kyutai) 0.88, LiveKit 0.74, Deepgram Flux 0.68, Pipecat 0.56.
• User pauses mid-thought: Causeur 0.92, Moshi 0.87, Pipecat 0.66, LiveKit 0.59, Deepgram Flux 0.57.
• User interrupts: Causeur 0.87, Moshi 0.67, LiveKit, Deepgram Flux and Pipecat 0.02 each.

The interruption row is the one I would watch. Most voice agents today either talk over you or stop dead at any noise, and the standard turn detectors in that chart barely register an interruption at all.

What does the Atmee V2 renderer do?

The renderer takes Causeur's decisions plus the voice audio and draws every frame: lip sync, micro-expression, head and body motion, and the scene around the character. Atmanity says it runs live on a single GPU, which is what makes the per-minute price possible. Because it is a general video foundation model that happens to be trained hard on people, the same renderer also produces offline clips, which is why it can be compared with Seedance at all.

Atmanity's avatar benchmark puts it against other real-time avatar models (metrics are Atmanity's own and not defined in detail):

• LipSync: Atmee V2 7.61, Runway 7.16, LemonSlice 6.74, LiveTalk 3.14, Kling AI 2.98.
• Naturalness: Atmee V2 0.741, Runway 0.629, LemonSlice 0.575, Kling AI 0.526, LiveTalk 0.503.
• Identity preservation: Atmee V2 0.881, LemonSlice 0.821, Runway 0.811, Kling AI 0.642, LiveTalk 0.543.
• Liveliness: Atmee V2 1.31, Runway 1.21, Kling AI 1.05, LemonSlice 0.96, LiveTalk 0.77.

What research is behind Causeur?

Atmanity has not published a Causeur paper yet; its site lists "a multimodal conversational behavior model" as forthcoming. Two papers the same team published earlier tackle exactly the problems Causeur claims to solve. To be clear, this link is my reading, not something Atmanity states:

• MM-When2Speak (arXiv 2505.14654, May 2025) watches a conversation's video, audio and transcript, encodes each separately, fuses them with self-attention and predicts one of three actions: stay silent, give a full response, or give a short reaction. F1 scores were 0.89, 0.84 and 0.72, against 0.71, 0.63 and 0.41 for a text-only LLM. That is the timing half of Causeur.
• GLARE (arXiv 2609.40317, NeurIPS 2026) turns speaker audio and one listener photo into a listener video that nods, smiles, laughs, frowns or looks surprised at the right moment. It is trained on about 147 hours of two-person conversations with 64,000+ labelled reactions, uses prosody (pitch, loudness, emphasis) to find the moments that deserve a reaction, and penalises mistimed reactions with a per-frame loss and a reward. On the RealTalk benchmark it beats DyStream, ListenFormer, ViCo, DIM and L2L on reaction F1 (0.59) and timing error (57.4 vs 63.2 for DyStream). That is the active-listening half.

There is also a data loop: Atmanity says atmee.ai, the consumer app, feeds back thousands of hours of real conversations every month for training.

Who makes Atmee V2?

Atmanity Inc., founded in October 2024 and based in Santa Clara. The CEO, Chen-Ping Yu, previously founded Phiar, an AR navigation startup that Google acquired in 2022. The team lists six PhDs, and the company names Forerunner, Essence, Recursive Ventures, Hyphen Capital, GFR Fund and Hawktail as backers. Atmee is the product name; Atmanity still shows up in the infrastructure, for example the API host api.atmanity.us and the Atmanity GitHub organisation.

Atmee V2 vs Seedance: which is faster and cheaper?

For short character clips, Atmee V2 is much faster and cheaper than Seedance in Atmanity's own side-by-side. The demo uses one prompt, "Selfie video of a blonde girl, stroking her hair", and renders a 10-second clip on both models:

• Seedance 2.5: 233.6 seconds, $0.50 per clip.
• Atmee V2: 3.5 seconds, under $0.01 per clip.

Watch the clip with sound

Atmanity's side-by-side from the launch video: same prompt, Seedance on the left (233.6 s, $0.50), Atmee V2 on the right (3.5 s, under $0.01). Play it with sound for the narration.

That clip is a 67x speed gap (233.6 / 3.5). A week before launch, the founder posted a second side-by-side, an 8-second clip:

• Seedance 2.5: 331.0 seconds, $1.19.
• Atmee V2: 6.4 seconds, $0.01.

Watch the clip with sound

The pre-launch comparison: Atmee V2 (6.4 s, $0.01) next to Seedance 2.5 (331.0 s, $1.19). Atmanity framed it as 51x faster and 119x cheaper.

Atmanity framed that one as "119x cheaper and 51x faster than Seedance 2.5". The headline claim at launch is more careful: "up to 50x faster and 10 to 100x cheaper" than general video models such as Seedance 2.5 and Wan 3.0. In both samples the output quality is close; Atmee V2 holds face and hair detail at least as well.

Warning: Commenters on the 8-second post asked a fair question: both panels perform almost the same motion, which two independent models rarely do from one prompt. Atmanity did not publish the exact workflow (same prompt and seed, or one clip used as reference for the other). Until someone reproduces it, treat the comparison as a vendor demo, not a benchmark.

The structural difference: Seedance is an async job that returns a file. Atmee V2 returns the file in seconds, or skips the file and streams into a call.

Why is Atmee V2 so much faster than Seedance?

Because it is built for a narrower job. Seedance is a general-purpose video generator: any scene, any camera move, many subjects, up to 30 seconds with generated sound in 2.5. You submit a job and poll for the file, which takes from about 30 seconds to a few minutes. Atmee V2 is specialised for one subject, a character facing the viewer, and it has to run at conversation speed on a single GPU. A model that only has to be good at people talking can spend far less compute per frame than one that has to draw anything.

Is Seedance better at anything?

Yes, and it matters if you are choosing between them:

• Scope. Seedance renders whole scenes, multi-shot sequences and non-character content. Atmee V2 is tuned for characters.
• Length. Atmee clips run 1 to 15 seconds (at 25 fps, 480x480 up to 1280x720; 1080p is listed as coming soon). Seedance 2.5 goes to 30 seconds.
• Access. Seedance is a public API through several providers. Atmee V2 is invite-only for now: building V2 avatars and clips needs the Growth plan or an invite code.
• Independent proof. The speed and cost comparison is Atmanity's own. No methodology was published, and I found no third-party benchmark yet.

Note: The "under $0.01" in the demo looks like a compute cost, not a list price. Atmee's published clip price at 480x480 is 0.3 credit for up to 10 seconds, and a credit costs about 10.5 to 15 US cents depending on the plan. That is roughly 3 to 4.5 cents for the demo's clip, still about ten times under the $0.50 Seedance figure.

Can Atmee V2 hold a live conversation?

Yes, and this is the part Seedance cannot do at all. Seedance and Wan return a finished file; neither can listen to you and answer while you watch. Atmee V2 streams video into a LiveKit room while Causeur handles turn-taking. Here is Atmanity's 36-second V2 reel of live calls:

Watch the clip with sound

Atmee V2 live sessions from atmee.ai: active listening, gestures and expression across different characters and roles.

Watch the listeners in that reel, not the speakers. The nods and small smiles while the user talks are Causeur at work; the drawing of them is the renderer.

How does the Atmee V2 API work?

There are three ways to use it, all on https://api.atmanity.us with an X-Api-Key header from your backend (the API refuses browser calls):

1. Full-stack session (POST /v1/sessions): Atmee runs speech recognition, Causeur, the LLM, the cloned voice and the video. Your page joins a LiveKit room and talks.
2. Avatar-only render (POST /v1/avatar_sessions, or the livekit-plugins-atmee Python package): you keep your own LiveKit voice agent and Atmee only draws the face.
3. Clips (POST /v1/generation/clips): image plus prompt, optional audio, mp4 out.

One full-stack session. Your backend never streams video; it only starts and ends the session.

Start a live session

Your backend starts the session and hands the page a one-room token. The call answers once the avatar is in the room, usually 15 to 20 seconds; the user then has 60 seconds to join.

terminal

$ curl -X POST https://api.atmanity.us/v1/sessions \ -H "X-Api-Key: $ATMEE_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "avatarId": "<avatar-id>", "maxDurationSeconds": 300 }' 

201 Created

{ "sessionId": "550e8400-e29b-41d4-a716-446655440000", "serverUrl": "wss://atmanity.livekit.cloud", "roomName": "rk-8f14e45f-ceea-467a-9f21-4c2f5b3a91cd", "userToken": "eyJhbGciOi...", "actualDurationSeconds": 300, "requestedDurationSeconds": 300, "durationCapped": false, "billingMode": "metered" } 

The page joins with livekit-client, attaches the avatar's tracks to a <video> and an <audio> element, and publishes the microphone:

avatar-room.js

import { Room, RoomEvent, Track } from "livekit-client"; const room = new Room(); room.on(RoomEvent.TrackSubscribed, (track) => { const el = document.getElementById(track.kind === Track.Kind.Video ? "avatar" : "voice"); track.attach(el); }); await room.connect(session.serverUrl, session.userToken); // within 60 seconds await room.startAudio(); await room.localParticipant.setMicrophoneEnabled(true); 

End it with POST /v1/sessions/{id}/end. Billing runs by the second, only while both you and the avatar are in the room.

Tell the avatar what is on screen

The most useful part for app builders is the atmee.app topic. The page sends small JSON messages and the avatar reacts in its own words: state replaces "what is true now", note adds a fact for the rest of the session, and event asks the avatar to remark on something once nobody has the floor.

app-messages.js

const send = (msg) => room.localParticipant.sendText(JSON.stringify(msg), { topic: "atmee.app" }); send({ type: "state", text: "The user is on the pricing page. Starter is 9 dollars a month, Growth is 49." }); send({ type: "note", text: "The user's name is Sara and she is on the free plan." }); send({ type: "event", text: "The user just finished checkout. Congratulate her in one sentence.", ttl_s: 12 }); 

Nothing comes back on that topic, so the only proof a message landed is what the avatar says next. One practical detail from the docs: wait about 2.5 seconds after the avatar arrives before the first message, or it is dropped.

Let the avatar call your app

App tools use the OpenAI function format, declared per session in appTools. Each call reaches the page as a LiveKit RPC named atmee.tool. notify tools let the avatar keep talking; await tools make it wait for your answer (up to 28 seconds).

appTools

{ "instructions": "Put each slide up before you teach it. When the trainee asks for their score, get it.", "tools": [ { "type": "function", "name": "show_slide", "description": "Put lesson slide n on the trainee's screen.", "parameters": { "type": "object", "properties": { "n": { "type": "integer" } }, "required": ["n"] } }, { "type": "function", "name": "get_score", "description": "Get the trainee's quiz score so far.", "parameters": { "type": "object", "properties": {} }, "x-atmee": { "mode": "await", "timeoutSeconds": 8 } } ] } 

tools.js

room.registerRpcMethod("atmee.tool", async ({ payload }) => { const { name, arguments: args } = JSON.parse(payload); if (name === "show_slide") { showSlide(args.n); return "ok"; } if (name === "get_score") return JSON.stringify(await fetch("/api/score").then((r) => r.json())); throw new Error(`no tool ${name}`); }); 

Register the handler before room.connect, so no call arrives at a page without one. Rules, scores and hidden answers belong in your app; the tool hands the avatar the result. Otherwise the avatar answers before your app acts, and may invent an outcome it later has to take back.

Give your existing LiveKit agent a face

If you already run a voice agent on LiveKit Agents, the plugin adds the avatar without moving your speech, LLM or tools:

agent.py

from livekit.plugins import atmee avatar = atmee.AvatarSession(avatar_id="...") # reads ATMEE_API_KEY await avatar.start(session, room=ctx.room) # before session.start await session.start(agent=MyAgent(), room=ctx.room) 

Your agent's TTS audio drives the face. Atmee never sees your LiveKit secret: you mint the avatar's token yourself. Version 0.2.0 and later render V2 avatars; a Pipecat plugin exists too.

Generate a clip

terminal

$ curl -X POST https://api.atmanity.us/v1/generation/clips \ -H "X-Api-Key: $ATMEE_API_KEY" \ -H "Idempotency-Key: 6f1c2a9e-3b7d-4c11-9a52-0d8e4f7b2c61" \ -H "Content-Type: application/json" \ -d '{"image":"https://example.com/host.png","prompt":"she holds up the bottle and smiles, the camera slowly pushing in","audio":"https://example.com/line.mp3","durationSeconds":12}' 

The response carries a short-lived mp4 link (24 hours). Without audio the clip is silent; with it, the character speaks the line. For longer videos, stitch 15-second clips with ffmpeg.

How much does Atmee V2 cost?

Everything is paid in credits. A credit costs about 10.5 to 15 US cents depending on the plan (prices read on 8 November 2026):

• Full-stack V2 session: 2 credits a minute, about $0.21 to $0.30 a minute. V1 is 1 credit.
• Avatar-only V2 render: 1 credit a minute. V1 is 0.5.
• Voice-only session: 1 credit a minute.
• Clips: 0.1 credit (480x480, up to 5 s) to 2 credits (1280x720, 15 s).
• Plans: Free (10 credits, 5-minute sessions, watermark), Starter $9, Growth $49 (needed to build V2 and clips), Pro $199, Scale $599, Enterprise for V2 HD.

So the "10 to 100x cheaper" claim is against general video models. Against other live-avatar vendors the price is in the same range: HeyGen LiveAvatar, Anam and Tavus all sit around $0.09 to $0.37 a minute. Atmee competes there on behavior and full-body motion, not on price.

What are the limits right now?

• Invite-only V2. Request a code at atmee.ai/v2signup, or be on Growth or higher. My own account is on the Free plan and reports v2Access: "invite_required", which is why the clips here are Atmanity's and not mine.
• Cold starts. V2 runs on a GPU pool that scales down when quiet. The first session after a quiet period can take minutes; the docs suggest a client timeout of seven minutes or a retry on 503.
• One conversation per avatar at a time, one user in it. More users need more avatars.
• No published latency number. The docs give 15 to 20 seconds to start a session, but no per-turn response time to compare with vendors who quote 150 to 600 ms.
• Claims, not audits. The 50x, the 10 to 100x and the one-GPU figures come from Atmanity. Treat them as a strong first data point until someone outside the company measures them.

Who should use Atmee V2?

Pick Atmee V2 when the product is a conversation with a face: a tutor, a sales role-play with a tough buyer, a museum guide, a support agent, a game character, a recruiter screening call. The agenda feature (a graph of conversation steps per avatar) and app tools make those shapes straightforward. Pick Seedance or a similar model when you need a cinematic shot, a scene with no person in it, or a clip longer than 15 seconds. And if you only need a quick clip of a person moving or speaking, Atmee V2 is the faster and cheaper option to try first.

Frequently asked questions

What is Atmee V2?

Atmee V2 is a real-time video model by Atmanity for AI avatars. From one image and about ten seconds of voice, it creates a character that can hold a live video call with expressions and gestures, or render short mp4 clips. It launched on 4 November 2026.

Is Atmee V2 better than Seedance?

For character clips and live avatars, yes on speed and cost: Atmanity's side-by-side shows a 10-second clip in 3.5 seconds for under $0.01 on Atmee V2, against 233.6 seconds and $0.50 on Seedance 2.5. Seedance is still the broader tool for full scenes, longer clips and non-character video.

What is Causeur in Atmee V2?

Causeur is Atmanity's behavior model, the "brain" of Atmee V2. It decides personality, emotion, timing and active listening moment to moment: when to speak, when to wait, when to nod or smile, and when to call an app tool. The Atmee V2 renderer, the "body", then draws those decisions as live video.

How is Atmee V2 different from Atmee V1?

V1 animates a talking head from the shoulders up at 512x512 and is cheaper (1 credit a minute). V2 renders the full frame, face, body, gestures and scene, in five aspect ratios, adds Causeur-driven emotion and timing, and costs 2 credits a minute. Anyone can talk to a V2 avatar; building one needs an invite or the Growth plan.

How fast is Atmee V2?

Atmanity claims up to 50x faster than general video models such as Seedance 2.5 and Wan 3.0, and fast enough to render a live call on one GPU. A live session takes about 15 to 20 seconds to start.

How much does Atmee V2 cost?

A full-stack V2 conversation costs 2 credits a minute, about $0.21 to $0.30. Avatar-only rendering is 1 credit a minute, and clips cost 0.1 to 2 credits each. Building V2 avatars needs the $49 Growth plan or an invite code.

Who makes Atmee?

Atmanity Inc., an AI lab in Santa Clara founded in 2024 by Chen-Ping Yu, who previously founded the AR startup Phiar, acquired by Google in 2022.

Can I use Atmee V2 with LiveKit?

Yes. Full-stack sessions run in a LiveKit room your page joins with a token, and the livekit-plugins-atmee Python package gives an existing LiveKit Agents voice agent an Atmee face. A Pipecat plugin is also available.

Originally published on my site. More posts and projects at mahmadmanzoor.com.