Generate
Back to Models
Models/Vidu/realtime avatarPass-through

Vidu S1

Vidu S1 turns one image and a persona into a continuously rendered character for natural voice conversations and instruction-driven performance. It supports photoreal people, fantasy characters, and pets, with audio-only or camera-aware video interaction; MachGen exposes the experience through a secure, metered LiveSession.

Current MachGen support
Realtime AvatarVoice-driven interactionCamera-aware video540p / 25 FPSPeople, characters & pets
Starting at $0.007 / output s
All supported sizes$0.007/s
Live public pricebookSee full pricebook
Overview

Realtime characters that listen, respond, and perform

Vidu S1 is a realtime interactive video model. It turns one character image and a persona into a continuously rendered conversation partner that can listen to the user, answer in natural speech, and follow spoken performance instructions.

The source image can depict a photoreal person, a fantasy or anime-style character, or a pet. Audio mode carries a two-way voice conversation; video mode also publishes the user's camera so the character can respond with visual context.

Capabilities

What Vidu S1 can do

  • Natural voice interaction. Hold a low-latency spoken conversation instead of waiting for a generated clip.
  • Voice-directed performance. Spoken instructions can guide the character's expression and movement during the session.
  • Camera-aware video calls. When the camera is published, S1 can use the user's visual reaction as additional interaction context.
  • Continuous realtime rendering. Vidu presents S1 as a continuously interactive API experience; its current Live API reference documents a 600-second maximum per session.
  • Realtime video output. Vidu publishes 540p at 25 FPS as the standard interactive output profile.
  • Flexible character styles. A single image can initialize realistic people, illustrated or fantasy characters, and pet mascots.
Character design

The image defines appearance; the persona defines behavior

A strong character setup covers four things: who the character is, the traits they should express, how they speak, and what they should help the user accomplish. Vidu recommends a focused persona of roughly 250 English words rather than a long, conflicting script.

For better motion and interaction

  • Use a clear, well-lit image with the face and intended body area visible.
  • Use a full-body source when spoken instructions need leg movement, stepping, jumping, or larger gestures.
  • Keep action requests direct. Complex multi-part physical instructions have a higher failure rate.
  • Describe dialogue style and response behavior explicitly, such as concise, patient, playful, or proactive.
  • Enable video mode only when the experience benefits from camera context; audio mode is the simpler voice-only path.
Input & output

One character image, two realtime channels

PartCurrent MachGen inputPurpose
Character imageOne public HTTPS image or base64 image data URI. The Playground accepts JPEG, PNG, or WebP up to 20 MB.Defines the character's visible identity and available body framing.
PersonaRequired, up to 2,000 characters; optional display name up to 128 characters.Defines role, personality, dialogue style, expertise, and response policy.
VoiceOptional avatar.voice identifier through the public API. The Playground currently uses the provider default voice.Selects how the character speaks. Voice cloning and voice-library management are not MachGen LiveSession endpoints today.
User mediaAudio mode publishes microphone audio. Video mode publishes microphone and camera.Carries the user's speech and, in video mode, visual context.
Character outputRealtime audio, plus video in video mode, delivered through AliRTC.Renders the speaking character continuously during the live session.
Use cases

Built for interactive roles, not prerecorded clips

  • Customer service and product guidance: answer pre-sales or after-sales questions and make conversational recommendations.
  • Education and training: explain concepts, check understanding, and guide a learner through the next step.
  • Culture and tourism: create an in-character guide aligned with a place, historical setting, or exhibition.
  • Personalized commerce: provide a one-to-one product presentation with a consistent host persona.
  • Game and community companions: keep a recognizable fantasy, anime-style, or mascot identity in an ongoing voice conversation.
On MachGen

What the public LiveSession API exposes today

MachGen supports custom images, personas, optional display names and voice identifiers, audio or video calls, and configurable session duration.

The current public contract does not expose Vidu's voice-cloning or voice-list APIs, text-message and manual-interrupt commands, transcription, VAD and LLM tuning, proactive-conversation timing, or external memory and knowledge retrieval. These vendor capabilities should not be treated as available MachGen controls yet.

MachGen accepts an even max_session_seconds from 30 through 600, matching the maximum currently documented in Vidu's Live API reference.

Official sources

Vidu S1 overviewVidu S1 API documentationVidu S1 introduction
Flow

One session, two channels

  1. Create the LiveSession over MachGen REST with the character image, persona, call mode, and maximum duration.
  2. Join AliRTC with the returned session-scoped credentials and publish microphone audio, plus camera video when requested.
  3. Connect to the protected MachGen control WebSocket and wait for the avatar to progress through connecting, preparing, and live.
  4. Interact after both the media and control channels are ready.
  5. Send the end command and leave AliRTC when the conversation finishes.

Audio and video stay on AliRTC. The MachGen WebSocket carries readiness and hangup signaling only.

Billing

Authorized first, settled after hangup

The selected maximum duration is reserved when the session is created. Billing begins when Vidu reports the avatar live, and MachGen refunds the unused authorization after final billed seconds arrive.