> ## Documentation Index
> Fetch the complete documentation index at: https://docs.avaluma.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Avatar Server Overview: Deploy Photorealistic AI Avatars

> Run the Avaluma Avatar Server to render .hvia avatars into live video and stream them into LiveKit rooms. Requires an NVIDIA GPU with 6+ GB VRAM.

The Avaluma Avatar Server is a Docker-based service that renders photorealistic AI avatars from `.hvia` files and streams the resulting video directly into a LiveKit room. It handles all GPU-accelerated rendering locally, giving you full control over compute resources and latency. You can run multiple avatar sessions at the same time — each session consumes approximately 2.5 GB of VRAM — making it straightforward to scale up as your needs grow.

## How It Works

The Avatar Server sits at the core of the Avaluma stack. A LiveKit Agent drives the conversational AI pipeline — speech-to-text, language model, and text-to-speech — and sends audio to the Avatar Server. The server animates the avatar in real time and streams the video back into the LiveKit room for participants to see.

## GPU Requirements

<Note>
  The Avatar Server requires a dedicated NVIDIA GPU. CPU-only environments are not supported.

  | Requirement              | Detail                                                                                                                |
  | ------------------------ | --------------------------------------------------------------------------------------------------------------------- |
  | CUDA version             | 12                                                                                                                    |
  | Additional capabilities  | OpenGL support, graphics drivers                                                                                      |
  | Minimum VRAM             | 6 GB (each avatar session uses \~2.5 GB)                                                                              |
  | Tested architectures     | Ampere, Ada Lovelace, Blackwell                                                                                       |
  | NVIDIA Container Toolkit | Required — [installation guide](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html) |
</Note>

## Deployment Options

You can connect your LiveKit Agent to either the Avaluma-hosted server or your own self-hosted instance:

* **Avaluma Hosted** — point your agent at `api.avaluma.ai` and let Avaluma manage the infrastructure.
* **Self-Hosted** — run the Docker service on your own hardware for full control over data and compute.

## Directory Structure

The `avatar-server/` directory contains everything you need to get started:

```
avatar-server/
├── assets/
│   └── avatars/          # Place your .hvia avatar files here
├── reverse_proxy/        # Optional: Caddy HTTPS reverse proxy
│   ├── Caddyfile
│   └── docker-compose.yaml
└── docker-compose.yaml
```

Place your `.hvia` avatar files in `assets/avatars/`. The `reverse_proxy/` folder contains an optional Caddy configuration for terminating TLS in production.

## Next Steps

<CardGroup cols={2}>
  <Card title="Setup" icon="server" href="/en/self-hosting/avatar-server/setup">
    Deploy the Avatar Server with Docker, add your avatar files, and configure the environment.
  </Card>

  <Card title="HTTPS Proxy" icon="lock" href="/en/self-hosting/avatar-server/https-proxy">
    Add automatic TLS to your deployment using the included Caddy reverse proxy.
  </Card>
</CardGroup>