Duix.Avatar: An Offline AI Avatar Toolkit That Runs on Your Own GPU

What Duix.Avatar actually solves, and for whom

The pitch in the README is cost. It compares the project's approach to traditional 3D digital human pipelines, which it describes as costing hundreds of thousands of dollars, and says the AI-generated route cut that to about $1,000. Whether or not that number holds in your case, the underlying claim is narrower and more useful: you can build a talking-head avatar from real-person video data rather than modelling a 3D character by hand.

The audience the README names is specific. It lists educators, content creators, legal experts, medical practitioners and entrepreneurs, and says the software is designed so that someone with no technical background can use it. That is consistent with the packaging: the repository is an Electron application with a Vue front end, not a Python library you import. There is a GUI, a model manager, and a one-click startup package for the Docker services.

The second half of the pitch is privacy. The README states that no internet connection is required and that this protects user privacy by avoiding data exposure during network transmission. For anyone producing video of a real person's face and voice, that is the deciding argument, not the price. A cloud avatar service needs your source footage uploaded; this one is documented as fully offline. If your material cannot leave the building, the offline constraint is the feature.

The architecture: an Electron shell over three Docker services

The repository layout makes the split visible. The top level holds `package.json`, `electron.vite.config.mjs`, `electron-builder.yml` and a `src/` directory, which is the desktop client. Alongside them sits `deploy/`, which the README says contains the `docker-compose.yml` file. The heavy work does not happen in Node.

Three images are named for the reader to pull. `guiji2025/fun-asr` handles automatic speech recognition, converting recorded speech into text. `guiji2025/fish-speech-ziming` handles the voice side, which the README describes as cloning a voice from samples and covering context, intonation and speed. `guiji2025/duix.avatar` is the avatar and video synthesis service, where facial features are captured and lip movement is matched to the audio.

The client's dependency list confirms its role. `axios` is present for HTTP calls to those local services, `better-sqlite3` for local storage of projects and models, `fluent-ffmpeg` for media handling, and `electron-updater` for application updates. The application is a controller and a project manager; the models run in containers. That is also why the hardware requirements are written as Docker and NVIDIA driver requirements rather than as npm requirements, and why the primary language of the repository is C.

Installing Duix.Avatar on Windows with Docker

The README documents two deployment modes, Windows and Ubuntu 22.04. The Windows path is the one spelled out in detail. It requires Windows 10 build 19042.1526 or higher, Node.js 18, an NVIDIA graphics card with drivers installed, and Docker. Disk space is the constraint people miss: the README asks for more than 30GB free on D for digital human and project data, and more than 100GB free on C for service image files, with a note that you can point Docker at another folder if C is short.

Start by checking that WSL is present, since Docker Desktop on Windows depends on it.

wsl --list --verbose
wsl --update

The first command lists installed distributions with their state; if it prints a distribution, the README says WSL is already installed and you can skip ahead. The second updates it. After that, install Docker for Windows from docker.com and start it.

With Docker running, pull the three images the README names. These are large, and this is the step that consumes the C drive space.

docker pull guiji2025/fun-asr
docker pull guiji2025/fish-speech-ziming
docker pull guiji2025/duix.avatar

Finally, bring up the services. The README states the `docker-compose.yml` file lives in the `/deploy` directory and that you run the following from inside it.

cd deploy
docker-compose up -d

The `-d` flag runs the stack detached, so the containers keep running in the background while you use the desktop application. The README also mentions a lite version of the compose file, but the excerpt cuts off before explaining how the lite variant differs, so treat that as something to check in the `deploy` directory itself.

For the desktop client, `package.json` defines the usual scripts. `npm run dev` starts the Electron app in watch mode for development, and `npm run build:win` produces a Windows build through electron-builder. If you only want the released application, the homepage at duix.com is where the project points for downloads.

What the requirements leave out

The README is unusually explicit about hardware and unusually quiet about everything else. There is no documented macOS or Linux desktop build. The scripts in `package.json` include `build:linux`, and the README offers an Ubuntu 22.04 deployment mode, but the system requirements section only states Windows 10 19042.1526 or higher, and the introduction calls the tool one designed for Windows systems. Anyone on a Mac should assume it will not work rather than assume the Linux path covers them.

The NVIDIA requirement is absolute as written. The recommended configuration is a 13th Gen Intel Core i5-13400F, 32GB of memory and an RTX 4070, with a note to ensure you have an NVIDIA card with drivers properly installed. There is no CPU-only fallback documented. On integrated graphics or an AMD card, the README gives no path forward.

The licence is the other gap. The repository's licence field reads NOASSERTION, and the root contains both a `LICENSE` file and a PDF titled "Duix.Avatar model community Licensing Agreement", in English and Chinese. That naming suggests the software and the models you clone with it may carry different terms, and the README excerpt does not resolve which applies to generated output. Read those files before building a commercial workflow on top of this.

Duix.Avatar compared with cloud avatar services

The obvious alternative is a hosted avatar service, and the README itself frames the comparison by naming the cost of traditional 3D digital human production. The real difference is not price per video but where the computation and the data live.

A hosted service takes your source video, runs the cloning on its own hardware, and returns rendered clips. You pay per render or per seat, you need a network connection every time, and the source footage of a real person's face and voice sits on someone else's infrastructure. Duix.Avatar inverts all three: the GPU is yours, the network is optional once the images are pulled, and the only copy of the model is on your disk. The cost moves from a subscription to a one-time hardware purchase and the electricity to run it.

That inversion is also the limitation. A cloud service scales to whatever queue you are willing to pay for. A local RTX 4070 renders one machine's worth of video, and the README's recommended configuration is a consumer card, not a datacentre accelerator. For a creator producing a handful of videos a week, local is cheaper and private. For a team producing hundreds, the cloud service wins on throughput even after you account for the data handling. The project's own positioning, aimed at individual professionals rather than production houses, matches that boundary.

Languages, maintenance and upgrade cost

The README states that scripts support eight languages: English, Japanese, Korean, Chinese, French, German, Arabic and Spanish. That is the text-to-speech input side. It does not claim the cloned voice reproduces all eight; the voice model is trained from your own samples, so its output quality depends on the samples you provide, not on a language list.

On maintenance, the facts are these. The repository is not archived, and the last push was on 2026-04-21. The most recent release listed is v1.0.6 from 2025-09-28, preceded by v1.0.5 on 2025-08-15. The desktop client depends on `electron-updater`, so the application can fetch new versions, but the three Docker images are pulled by tag with no version pinned in the commands the README gives, which means a fresh `docker pull` can bring a newer service than the client was tested against. If you need reproducibility, pin the image digests yourself; the README does not.

Upgrade cost is mostly disk and bandwidth. Each new image pull consumes space on the drive Docker uses, and the README already asks for more than 100GB there. A machine that barely met the requirement at install time will need attention after a few releases. The Node.js side is lighter, but the project pins Node.js 18 in its dependencies section, so an upgrade to a newer Node major is an untested change until the project says otherwise.

Editorial conclusion

Duix.Avatar fits people who already own an NVIDIA GPU and Windows 10 19042.1526 or newer, have roughly 130GB of free disk across C and D, and care that the footage never leaves the machine. It does not fit macOS users, anyone without a discrete NVIDIA card, or teams that need a hosted API with per-request billing, because the README documents no such service. Before committing, verify three things on your own hardware: that `docker pull guiji2025/duix.avatar` completes, that your C drive really has 100GB free, and that the model community licence PDF in the repository root covers the commercial use you have in mind.

Frequently asked questions

Official sources

1. duixcom/Duix-Avatar on GitHub
2. Issues
3. Project website
4. README
5. Releases