Vitoom

English | 中文 | 日本語

Vitoom is a locally deployable AIGC application platform. Access it through a browser to run text, image, audio, and video inference on your own machine (personal PCs with DGX Spark, RTX Spark, or RTX 30/40/50 series GPUs). A built-in AI Agent orchestrates writing, translation, document processing, knowledge-base retrieval, and multimodal generation. It suits solo creators and small teams on a LAN. If your work demands strict privacy—you cannot expose data or rely on cloud LLMs—Vitoom is built for that use case.

Main use cases

Area	Description
Writing & office	Documents, reports, summaries; copy ideation; export conversations to Markdown / PDF
Knowledge base	Archive PDF, Word, PPT, etc.; semantic search and Q&A; build a private knowledge base over time
Voice & audio	Text-to-speech (multi-speaker, voice design, cloning); multi-character dialogue / drama-style dubbing; speech-to-text
Image & video	Text-to-image (mainstream open models), image editing; image Q&A; text-to-video / image-to-video
Documents & OCR	Summarize and convert web / PDF / Office links; scan OCR (tables, formulas); export tables to Excel
Translation	Long-form multilingual translation; text-in-image translation
Web search	Optional live web search (requires a Tavily API key)

Requirements

• Docker and Docker Compose (docker compose subcommand)
• Inference: NVIDIA GPU, NVIDIA driver with CUDA 13.0 support (matches cu130 inference images; check with nvidia-smi), and NVIDIA Container Toolkit (native Linux, or Docker Desktop + WSL2 on Windows)
• Python 3.10+ to run scripts/ setup and model download only (no full local inference stack required on the host)
• Network access to image/model sources (setup wizard Mainland China prefers domestic mirrors and ModelScope; Other regions mainly use Docker Hub / Hugging Face)

Platforms

Platform	Notes
Linux	Native Docker recommended; run commands from the repository root
Windows	Docker Desktop + WSL2; enable GPU and File Sharing for the project disk; run Python scripts and docker compose in the same environment (WSL2 terminal or PowerShell throughout—do not mix paths)

Optional GPU / CUDA 13.0 runtime check (Docker installed and GPU passthrough working):

docker run --rm --gpus all nvidia/cuda:13.0.0-base-ubuntu24.04 nvidia-smi

Windows Preparation

Windows users should complete the steps below first. Run the preparation commands in PowerShell. For the later installation steps, keep using the same environment: do not switch back and forth between PowerShell and a WSL terminal, or paths may not match.

1. Open PowerShell

Search for PowerShell in the Start menu and open it. First check that WSL is available:

wsl --version
wsl -l -v

In the wsl -l -v output, check the VERSION column. The Linux distribution you use must be version 2. If it shows 1, switch it to WSL2:

wsl --set-default-version 2
wsl --set-version <distribution-name> 2

Replace <distribution-name> with the name shown by wsl -l -v, for example Ubuntu.

2. Install Git

Git is required to download the Vitoom source code. Without Git, the later git clone command will fail.

winget install --id Git.Git -e --source winget

3. Install Python 3.11

Python is required to run the setup wizard and download scripts under scripts/.

winget install --id Python.Python.3.11 -e --source winget

4. Reopen PowerShell

After installing Git and Python, close the current PowerShell window and open a new one. Then verify the installation:

git --version
py -3 --version
docker compose version

If all commands print version information, continue with “Quick install”.

Quick install

First download the project code, then enter the project directory and run the install commands. Windows users should continue in the newly opened PowerShell window.

1. Clone the project code

git clone https://github.com/tonera/vitoom.git
cd vitoom

2. Configure environment

The setup wizard writes .env, detects x86_64 / aarch64, and sets LAN URLs for inference. During configuration, note: do not set VITOOM_BACKEND_URL to 127.0.0.1, or inference containers cannot reach the backend.

python scripts/setup_vitoom.py

On Windows PowerShell, if python is not found, replace later commands that start with python with py -3, for example:

py -3 scripts/setup_vitoom.py

3. Load images

python scripts/load_vitoom_images.py

Loads offline tar from images/<arch>/ when present, otherwise pulls from Docker Hub. Partial components example:

python scripts/load_vitoom_images.py --components backend,visual,text

4. Start services

Start backend first (creates Docker network vitoom-net), then inference:

docker compose up -d backend

Start inference profiles selected in the wizard (full stack below—one line, because Windows CMD does not support \ line continuation):

docker compose -f docker-compose.inference.release.yml --profile visual --profile text --profile audio --profile mini --profile download up -d

Subset example (image + text only):

docker compose -f docker-compose.inference.release.yml --profile visual --profile text up -d

Open in browser: http://<LAN-IP>:8888 (see VITOOM_BACKEND_URL / VITOOM_SERVER_PORT in .env; you may use 127.0.0.1 in the browser locally, but .env should still use the LAN IP).

5. Download models (optional, large)

python scripts/download_initial_models.py

Or download later under Models in the Web UI (requires the download profile). For a first try, at least start backend + text and download an LLM.

More detail: docker-usage-en.md (中文 / 日本語).

Usage

1. Sign in: Open http://<LAN-IP>:8888 in a browser; the default admin after first deploy is admin@vitoom.ai with initial password admin123456 (same as DEFAULT_ADMIN_PASSWORD in .env). Change this password immediately after first login. Administrators can add more users in the Web admin UI.
2. Agent: Chat in natural language for writing, translation, documents, knowledge base, image/audio/video generation (tools are chosen automatically).
3. Workspaces: Home → Image, Video, Audio (ASR/TTS), Translate, etc.
4. Models: Download and activate weights in the model list; needs the download profile or step 5 script.
5. Knowledge base: Archive files or conversations, then query via Agent.
6. Web search (optional): Set TAVILY_API_KEY in .env (Tavily).

First inference startup can be slow (loading weights). Logs:

docker compose logs -f backend
docker compose -f docker-compose.inference.release.yml logs -f visual

Related docs

Doc	Description
docker-usage-en.md	Docker deployment, profiles, data dirs, troubleshooting
docker-usage-jp.md	Same guide in Japanese
docker-usage-cn.md	Same guide in Chinese

Acknowledgments

• TurboDiffusion — fast video inference
• Nunchaku — image inference acceleration
• Qwen3-TTS — speech synthesis
• Qwen3-ASR — speech recognition
• VoxCPM — fast speech synthesis
• vLLM — efficient text inference
• RMBG-2.0 — background removal
• MeanCache — image inference acceleration

License

This project is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0). Commercial licensing: COMMERCIAL_LICENSE.md (if applicable).

Upstream models and third-party components have their own licenses—verify compliance before use.