Vidu S2-Avatar overview

Vidu S2-Avatar turns a photo of one person into a character you can talk to. You can define its role and behavior and optionally choose a voice. It responds with synchronized video and speech. It accepts voice or text input and can use the caller’s camera during a video call. Avatars are saved for up to 90 days and can be reused across sessions. Each session holds one avatar and one active call at a time.

Billing starts when the session reaches ready, including time between calls. Disconnect when the caller is done. See Billing for how session time maps to rates.

At a glance

Spec	Value
Model name	Vidu S2-Avatar
Pricing
Input	One photo, a character description, and voice or text; optional camera video
Output	Synchronized character video and speech
Frame rate	25 fps
Resolution	864 × 1152
Audio	48 kHz mono, in sync with the video

What you can do

Create and reuse an avatar

Create an avatar from one image. Save its avatar_id and attach it again in later sessions.

Talk or type

Publish a microphone or send text with say. Read transcript messages to show both sides of the conversation.

Steer a live call

Use interrupt to stop speech. Use update_call to change the persona or voice. Send reference images to change what the character holds, wears, or stands in front of.

Quick start

These examples use the base SDK to create an avatar and start a call. The tutorial shows how to publish a microphone and build the rest of the call UI.

import { Reactor } from "@reactor-team/js-sdk";

async function startAvatarCall(jwtToken, file, video, audio) {
  const reactor = new Reactor({
    modelName: "reactor/vidu-s2-avatar",
    modelTracks: [
      { name: "mic", kind: "audio", direction: "sendonly" },
      { name: "webcam", kind: "video", direction: "sendonly" },
      { name: "main_video", kind: "video", direction: "recvonly" },
      { name: "main_audio", kind: "audio", direction: "recvonly" },
    ],
  });

  reactor.on("trackReceived", (name, track) => {
    if (name === "main_video") video.srcObject = new MediaStream([track]);
    if (name === "main_audio") audio.srcObject = new MediaStream([track]);
  });

  function waitForPhase(phase) {
    return new Promise((resolve) => {
      const onMessage = (msg) => {
        if (msg.type !== "session_state" || msg.data.phase !== phase) return;
        reactor.off("message", onMessage);
        resolve(msg.data);
      };
      reactor.on("message", onMessage);
    });
  }

  await reactor.connect(jwtToken);
  await reactor.resumeTrack("main_video");
  await reactor.resumeTrack("main_audio");

  const avatarReady = waitForPhase("avatar_ready");
  const image = await reactor.uploadFile(file);
  await reactor.sendCommand("create_avatar", { image, name: "Tina" });
  await avatarReady;

  const live = waitForPhase("live");
  await reactor.sendCommand("start_call", {
    persona: "You are Tina, a museum guide. Answer in two short sentences.",
    greeting: "Say hello and wave.",
  });
  await live;
  await reactor.resumeTrack("main_video");
  await reactor.resumeTrack("main_audio");

  return reactor;
}

import asyncio
from reactor_sdk import Reactor

async with Reactor(model_name="reactor/vidu-s2-avatar", api_key="rk_...") as reactor:
    phases = {"avatar_ready": asyncio.Event(), "live": asyncio.Event()}

    @reactor.on_message
    def on_message(msg):
        if msg["type"] == "session_state" and msg["data"]["phase"] in phases:
            phases[msg["data"]["phase"]].set()

    await reactor.connect()
    await reactor.send_command("create_avatar", {"image_url": "https://YOUR_PUBLIC_HOST/tina.jpg"})
    await phases["avatar_ready"].wait()

    await reactor.send_command("start_call", {
        "persona": "You are Tina, a museum guide. Answer in two short sentences.",
    })
    await phases["live"].wait()
    await reactor.send_command("say", {"text": "What is in this room?"})

In a browser, start startAvatarCall from a click or tap so the browser can play audio.

How a call works

Connect. The first session_state has phase idle. All four tracks exist from connect, so the app can prepare its player and microphone before a call starts.
Create or reuse an avatar. Send create_avatar with an upload or a public image URL. The phase changes to preparing_avatar, then avatar_ready. Save the avatar_id. In a later session, send attach_avatar with that id instead.
Start a call. Send start_call with a persona. You can choose a voice from list_voices or use the default. The phase changes through starting and warming_up to live.
Talk and steer. Publish mic or send say. Use interrupt and update_call to steer a live conversation. Use set_reference_images to change what the avatar wears, holds, or stands in front of.
End or start again. Send end_call, or let the call end at its time limit or idle timeout. The avatar stays available, so the app can start another call in the same session.

Use session_state to drive the UI. The snapshot reports the phase, current avatar, call timers, and media state. A refused command or failure arrives as command_error with a stable code. See the schema for commands, messages, and tracks. The prompt guide covers personas, greetings, and turn-taking.

Overview

Vidu S2-Avatar

H3 Reference Turbo Realtime

FastH3

Visko Orbis Stable

Visko Orbis Dynamic

HappyOyster

LingBot World 2

LingBot

X2

Helios

LongLive-2.0

SANA-Streaming

LTX

Vidu S2-Avatar overview

At a glance

What you can do

Create and reuse an avatar

Talk or type

Steer a live call

Quick start

How a call works

​At a glance

​What you can do

Create and reuse an avatar

Talk or type

Steer a live call

​Quick start

​How a call works

At a glance

What you can do

Quick start

How a call works