Learn how LiveAvatar works, how it differs from HeyGen’s former Interactive Avatar, and what to know about avatars, modes, memory, and credits.
LiveAvatar is HeyGen’s platform for building real-time, conversational AI experiences with lifelike digital avatars. Introduced to the HeyGen Community as the next evolution of Interactive Avatar, it combines synchronized video with conversational AI so users can speak with an avatar in real time.
The platform supports ready-made and custom avatars, managed or bring-your-own AI infrastructure, embeddable experiences, and API-based integrations. Developers can begin with the LiveAvatar website or go directly to the LiveAvatar documentation.
What is LiveAvatar?
LiveAvatar provides programmable interfaces for adding a human-like visual presence to an AI agent. During a session, it renders a real-time, lip-synced video stream of an avatar that speaks and reacts as the conversation progresses.
A LiveAvatar session can process user input, pass it through a language model, generate speech, and render synchronized video. Depending on the selected mode, LiveAvatar can manage the full conversational pipeline or operate only as the real-time video layer.
The platform is designed for applications such as:
- Virtual sales assistants and product demonstrations
- Support and training agents
- Interactive tutors
- Digital hosts and characters
- Conversational experiences embedded in websites or products
These are documented use cases, not guarantees about the suitability or performance of a particular implementation.
How LiveAvatar differs from Interactive Avatar
HeyGen announced in its Community on November 6, 2025, that Interactive Avatar was becoming LiveAvatar. The announcement described LiveAvatar as a platform for hyper-realistic, real-time conversations with AI.
The current product has its own website, dashboard, API documentation, SDK, pricing structure, and developer workflows. The supplied Community discussion indicates that some Interactive Avatar assets, including at least one user’s knowledge base, were migrated during the transition. However, the present official documentation does not provide enough public information to guarantee migration behavior for every account or asset.
Anyone relying on an existing Interactive Avatar configuration should verify its status in the LiveAvatar dashboard rather than assuming that every avatar, knowledge base, voice, or integration was transferred automatically.
LiveAvatar integration modes
LiveAvatar offers two primary integration models: FULL Mode and LITE Mode.
FULL Mode
FULL Mode manages the main components of the real-time conversational pipeline:
- Voice activity detection
- Speech-to-text processing
- Large language model responses
- Text-to-speech generation
- WebRTC orchestration and video delivery
Developers configure the avatar, voice, conversational context, and session settings. LiveAvatar then coordinates the underlying real-time infrastructure. This mode is intended for teams that do not want to construct and maintain each part of the audio, AI, and streaming pipeline independently.
FULL Mode also permits custom components in supported configurations. The LiveAvatar documentation states that developers can connect an OpenAI-compatible language model or their own inference endpoint and use external text-to-speech providers.
LITE Mode
LITE Mode focuses on real-time avatar video generation. The developer supplies the speech-to-text, language model, text-to-speech, and conversation logic, while LiveAvatar turns the resulting audio into synchronized avatar video.
This approach is intended for teams that already have an AI or voice-agent stack, or that need greater control over conversation flow and response logic. The documentation identifies integration paths for systems including LiveKit, Pipecat, Agora, ElevenLabs Agents, OpenAI Realtime, and Gemini Live.
In short, FULL Mode manages the conversational stack, while LITE Mode lets developers bring their own stack and use LiveAvatar primarily as the visual layer.
LiveAvatar avatar options
LiveAvatar supports stock avatars and custom avatars.
Stock avatars
A library of stock avatars is available for projects that do not require a unique identity. Stock avatars do not consume custom avatar slots.
Some stock avatars are designed for background replacement. They render against a solid green background, allowing developers to apply chroma-key processing and place the avatar over a custom interface or branded backdrop.
Custom video avatars
A custom video avatar is created from recorded footage. According to the current documentation, the process requires:
- Two minutes of input footage.
- A separate consent recording from the person whose likeness is being used.
The training footage is used to model the person’s appearance, expressions, and mannerisms. Because the footage contains speech, the process also creates a voice clone, although another available voice can be selected when a session starts.
Custom image avatars
LiveAvatar also supports custom avatars created from a single image. This answers an unresolved question raised in the original Community discussion: the current documentation confirms that an image-based person can be used to create a LiveAvatar.
Unlike a video avatar, an image avatar does not generate a voice clone. A voice must be supplied from the available voice library, through a custom text-to-speech integration, or through the developer’s audio pipeline in LITE Mode.
Custom avatar creation is described as a paid subscription add-on and requires an available avatar slot. Both video and image avatars use slots; stock avatars do not.
Memory and continuing conversations
LiveAvatar provides an opt-in memory system for FULL Mode. By default, individual sessions are isolated. A new session does not automatically know what happened in earlier conversations.
When sessions are explicitly linked to the same memory, LiveAvatar can summarize key information from previous session transcripts and add that summary to a later session’s context. The memory is updated as additional linked sessions end.
This memory is not a searchable archive of every previous statement. The current documentation describes it as a single accumulated summary. It can be read or deleted through the API, but its contents cannot be manually edited. LITE Mode users must manage conversation history through their own AI stack.
Credits and session billing
LiveAvatar deducts credits according to session duration and mode:
- Embed and FULL Mode use two credits per minute.
- LITE Mode uses one credit per minute.
A session requires enough credits for at least one minute before it can begin. When overage billing is disabled, active sessions end after the available credit balance is exhausted, and new sessions cannot begin. Plan prices, allowances, session limits, concurrency, and overage terms can change, so readers should consult the LiveAvatar pricing page before making a purchasing decision.
Frequently asked questions
Is LiveAvatar the replacement for Interactive Avatar?
HeyGen’s November 2025 Community announcement presented LiveAvatar as the new identity and direction for Interactive Avatar. The current LiveAvatar product now operates through its own website, dashboard, and documentation.
Can I create a LiveAvatar from an AI-generated person?
The current documentation supports creating an image avatar from a single image. It does not require the two-minute recording used for video avatars. Users remain responsible for following HeyGen’s applicable policies and ensuring they have the necessary rights to submitted material.
Does a custom video avatar require consent?
Yes. LiveAvatar’s documentation requires a consent recording for a custom video avatar. The main training input is a two-minute video.
Can LiveAvatar remember earlier conversations?
Yes, in FULL Mode when sessions are explicitly linked to a shared memory. Memory is opt-in and is not applied automatically to unrelated sessions. LITE Mode integrations manage memory independently.
Do I need to build the entire AI stack?
No. FULL Mode can manage speech recognition, language-model responses, speech generation, and real-time streaming. LITE Mode is available when you want to provide those components yourself.
