Build Your Private AI Image Generator with Docker and Open WebUI

Introduction

Ever found yourself needing a quick image for a project, only to get sidetracked by credit counts, privacy worries, or overly strict content filters blocking your dragon-in-a-suit concept? There's a better way: run everything on your own machine with a polished chat interface. Docker Model Runner now makes it trivial to pull a diffusion model, connect it to Open WebUI, and start generating images—all locally, privately, and without a single subscription fee. This guide walks you through the entire setup, from prerequisites to your first generated image.

What You'll Need

• Docker Desktop (macOS) or Docker Engine (Linux)
• ~8 GB of free RAM for a small model (more RAM allows larger models)
• GPU (optional but highly recommended): NVIDIA CUDA, Apple Silicon (MPS), or CPU fallback
• Open WebUI (automatically handled by Docker Model Runner)

If you can run docker model version without errors, you're ready to proceed.

Step 1: Pull an Image Generation Model

Docker Model Runner uses DDUF (Diffusers Unified Format) to distribute models via Docker Hub. This single-file bundle contains everything a diffusion model needs: text encoder, VAE, UNet/DiT, and scheduler configuration. To get started, open your terminal and run:

docker model pull stable-diffusion

This pulls the default Stable Diffusion XL model. You can verify it downloaded correctly with:

docker model inspect stable-diffusion

You'll see output like this (truncated for clarity):

{
  "id": "sha256:5f60862074a4c585126288d08555e5ad9ef65044bf490ff3a64855fc84d06823",
  "tags": ["docker.io/ai/stable-diffusion:latest"],
  "config": {
    "format": "diffusers",
    "size": "6.94GB",
    ...
  }
}

The model is stored locally as a DDUF file. At runtime, Docker Model Runner unpacks it and starts the inference backend. For a list of available models, run docker model search.

Step 2: Launch Open WebUI

Here's the magic: Docker Model Runner includes a built-in command that wires up Open WebUI against your local inference endpoint. Just run:

docker model launch openwebui

This command automatically:

• Starts the model inference service (exposing an OpenAI-compatible API)
• Pulls and runs the Open WebUI container
• Connects the UI to the API endpoint
• Opens a browser tab at http://localhost:8080 (or your configured port)

You'll see a familiar chat interface. The backend fully supports the POST /v1/images/generations endpoint, so Open WebUI can use it natively.

Step 3: Generate Your First Image

Once Open WebUI loads in your browser:

1. Select the image generation model from the dropdown (usually named "stable-diffusion" or something similar).
2. Type a prompt in the chat box, e.g., "a dragon wearing a business suit, digital art".
3. Hit Enter and wait a few seconds. The model will generate an image and show it inline.
4. Refine your prompt or adjust parameters (like number of images, size) in the settings panel if available.

Because everything is local, there are no credit limits, no content filters (unless you add them), and no data leaving your machine.

Tips for Best Results

• Use a GPU for speed. CPU generation can be 10–100x slower. For NVIDIA, ensure CUDA drivers are installed. For Apple Silicon, Docker Desktop handles MPS automatically.
• Free up RAM. Stable Diffusion XL requires about 8 GB VRAM. Close other heavy applications to avoid swap thrashing.
• Experiment with models. Try docker model pull sdxl-turbo for faster generation or docker model pull pixart-sigma for higher quality. Use docker model search to discover more.
• Customize Open WebUI. You can add external Ollama models for chat while keeping image generation local. Just configure the OpenAI endpoint in Open WebUI settings.
• Manage storage. Image models are large (5–10 GB each). Run docker model prune to remove unused models.
• Security. Keep the Docker port unused or behind a firewall—local AI means local only.

That's it! You now have a private, no-cloud image generator accessible through a friendly chat interface. No subscriptions, no data leaks, no rejected prompts.