VirtualMan · Open-source Digital Human Stack

Free, high-quality, pluggable digital humans.

An open-source stack with Agent Skill, Provider API, and a default gateway for one shared digital-human workflow across agent environments.

View repository Quick start

Open-source and free High-quality rendering Pluggable providers Full digital-human stack End-to-end workflow

Digital-human demo video Default version

Workflow

Text mode and audio mode share one generation flow.

Use a reference video with either a script or continuous speech audio. The client submits the job, polls status, and downloads the result.

job lifecycle Reference video + text / audio → MP4 + record

Inputs

Script driven

Reference video + script

Write the script and make the person in the reference video speak new content.

--video + --text

Audio driven

Reference video + speech audio

Use existing speech audio and sync the reference person to that track.

--video + --audio

Processing

01

Submit job

The Skill packages input media into a request, submits it to the Provider API, and records the job ID.
02

Wait for result

The client keeps polling until the backend returns a downloadable video file.
03

Validate file

Validate decoding, audio and video streams, duration, and SHA-256.

Outputs

MP4 + task record

Outputs an MP4 video and saves a task record automatically. Most users only need the video; the record is for troubleshooting.

talking-head.mp4 task record

Architecture

Decouple agents from digital-human runtimes through the Provider API.

The Skill calls a stable HTTP contract. duix-gateway is the default implementation, and any compatible digital-human backend can replace it.

Interface map Agent orchestration and API live above gateway and rendering services.

Agent side The Skill calls the backend through an API and does not depend on a specific agent or runtime

Default backend The runtime can be replaced while the agent call pattern stays stable

Default implementation

Core capabilities

Agent orchestration, Provider API, and a default gateway.

VirtualMan does not include model weights. Its core is the HTTP API that connects Agent Skills with replaceable digital-human runtimes.

A Agent SkillStart talking-head generation inside agent environments that support Skills.

B Provider APISubmit jobs, check status, and download results through one contract.

C duix-gatewayDefault gateway with auth, port checks, and Docker Compose deployment.

D Task recordEach render keeps job ID, input hashes, and validation results for troubleshooting.

Quick start

Render your first video in three steps.

Install the Skill from npm for Codex, Claude, or Gemini. The default backend is best run on Linux or WSL with NVIDIA GPU, Docker, NVIDIA Container Toolkit, and FFmpeg.

01

Install the Agent Skill

Run the npm installer. Replace codex with claude, gemini, or all if needed.

02

Start duix-gateway

On a GPU machine, copy .env, set GATEWAY_TOKEN, and start the service.

03

Render video

Pass a reference video plus script or audio. The client waits for the backend, then downloads the MP4.

Open README View Provider API

3 commands · first video

First render

ready

01 Install

npx -y @infranic-org/virtualman-skill install codex

This example installs into Codex. claude, gemini, and all are also supported.

02 Start gateway

cp .env.example .env && ./scripts/start.sh

Set GATEWAY_TOKEN first. The start script checks ports before launch.

03 Render

python3 skills/create-digital-human-video/scripts/render.py --video reference.mp4 --text "Hello, this is a VirtualMan talking-head test."

Wait for rendering and download the MP4. The task record is saved next to the result automatically.

Start with one render

Render a text-driven talking-head video from one reference clip.

Once the end-to-end path is verified, plug in your own backend, media rules, and validation policy.

Star the project Quick start