Open-source 3D digital human engine
A real-time, audio-driven digital human framework built on dual render backends. Open-source core with enterprise-grade deployment. From a single portrait to a speaking, listening, thinking avatar.
Send the free Studio package to a customer server. They run two scripts, open the web console, upload an avatar, and start a training job.
Audio drives lip movement frame-by-frame. Low-latency streaming pipeline turns speech into a talking avatar in under two seconds — ask a question, get a face that answers.
Two render backends under one interface: a 2D portrait engine for real-time conversation, and a full-body 3D engine for cinematic, offline-quality output. Switch without touching business logic.
The full engine is open-source. Build on LHM and LivePortrait with a clean abstraction layer. No black boxes — every render step is inspectable and replaceable.
Self-hosted, LAN-deployable, brand-consistent. Custom avatars in your uniform, domain-specific knowledge bases, and full control over data. Yours to own, not rent.
A streaming pipeline chains speech recognition, reasoning, synthesis, and rendering. Each stage runs as soon as its input arrives, so the avatar starts speaking before the full answer is generated.
TTS audio is split into frames and mapped to lip-open ratios, fed directly into the portrait renderer. No driving video needed — the voice animates the mouth.
A portrait is preprocessed once — face detection, 3D feature extraction, keypoints — then reused across unlimited conversations. Instant replies, no repeated compute.
For cinematic output, the 3D engine reconstructs an animatable human from a single image and drives it with motion data. Build the avatar once, reuse every motion.
Four ready-made personas — explainer, consultant, streamer, guide — each with its own system prompt, render backend, and motion profile. Switch scenes in config.
Plug in a domain knowledge base to ground the LLM. Real estate, finance, healthcare — the avatar answers with your facts, not hallucinations.
One renderer interface, two engines. Business code never knows whether a 2D portrait or a 3D avatar is behind the screen. Add backends without rewriting logic.
$ git clone OliveWolf
$ python app_realtime.py
↗ http://localhost:7860
Workspace
Avatar
Knowledge
Conversation
Publish
Private Deployment
Custom Avatar
Knowledge Base
Monitoring
Support
Clone the open-source engine, or bring OliveWolf Enterprise to your organization with custom avatars and on-premise deployment.