Your Private AI Image Studio: Running Stable Diffusion Locally with Docker and Open WebUI

Have you ever wanted to generate stunning AI art without worrying about privacy, credit limits, or overly strict content filters? Docker Model Runner now makes it possible to run an entire image generation pipeline on your own machine, with a clean chat interface provided by Open WebUI. No cloud subscriptions, no data leaving your computer. In this Q&A, we'll cover everything you need to know to set up your own private image generation studio — from hardware requirements to pulling models to launching the interface. Use the links below to jump to a specific question, or read straight through to get the full picture.

1. What is Docker Model Runner and how does it enable local image generation?

Docker Model Runner is a command-line tool that acts as a control plane for AI models. Instead of relying on cloud services, it downloads and manages models on your local machine, then exposes an OpenAI-compatible API — including the /v1/images/generations endpoint. This means any tool that talks to OpenAI’s image API can work with your local setup. In this guide, Docker Model Runner pulls a Stable Diffusion model, handles the inference backend lifecycle, and connects it to Open WebUI, a web-based chat interface. The result: you type a prompt in a familiar chat window, and the image is generated locally, fully private, with no data ever leaving your system.

Your Private AI Image Studio: Running Stable Diffusion Locally with Docker and Open WebUI — Source: www.docker.com

2. What hardware and software do I need to get started?

The requirements are modest but flexible. You need Docker Desktop (macOS) or Docker Engine (Linux) installed and running. For memory, plan on at least ~8 GB of free RAM for a small model — more is better if you want faster generation or larger images. A GPU is optional but highly recommended: NVIDIA (CUDA) on Linux, or Apple Silicon (MPS) on macOS. If you don’t have a compatible GPU, the CPU fallback will work but be noticeably slower. To confirm your Docker setup is ready, run docker model version — if you see version info without errors, you’re good to go.

3. How do I pull an image generation model using Docker Model Runner?

Once Docker is ready, pulling a model is a single command: docker model pull stable-diffusion. This downloads the model from Docker Hub, where it’s distributed as a DDUF (Diffusers Unified Format) package — a compact, single-file artifact that bundles all the components of a diffusion model (text encoder, VAE, UNet/DiT, scheduler config). After the download finishes, you can inspect the model with docker model inspect stable-diffusion. The output will show the model’s ID, tags, creation date, and configuration details, including the internal .dduf file name and size (e.g., 6.94 GB for the FP16 variant). This confirms the model is stored locally and ready to use.

4. What exactly is a DDUF file and why does it matter?

DDUF stands for Diffusers Unified Format. It’s a packaging format designed to simplify distribution of diffusion models. Instead of dealing with separate folders for the text encoder, VAE, UNet/DiT, and scheduler configuration files, DDUF bundles them all into a single file that can be stored as an OCI artifact on Docker Hub. At runtime, Docker Model Runner unpacks the DDUF file transparently. This matters because it makes model management as easy as pulling a container image — no manual downloading of multiple components, no version mismatches. You get a reproducible, portable artifact that’s ready to run on any compatible system.

5. How do I launch Open WebUI with Docker Model Runner?

This is the “magic trick” of the setup. Docker Model Runner includes a built-in launch command that automatically wires up Open WebUI against the local inference endpoint. Simply run: docker model launch openwebui. That’s it — no manual configuration, no exposing ports, no setting environment variables. The command starts Open WebUI in a way that it already knows how to talk to the Stable Diffusion model running locally. You’ll be greeted by a chat interface where you can type your prompt (e.g., “a dragon wearing a business suit”) and have the image generated right on your machine, with full privacy and no usage limits.

Source: www.docker.com

6. Is a GPU necessary, and how do I choose one?

A GPU is not strictly necessary but it makes a huge difference. If you have an NVIDIA GPU with CUDA support (common on Linux) or an Apple Silicon chip (M1, M2, M3) on macOS, Docker Model Runner will automatically use the hardware acceleration via MPS. With a GPU, image generation typically takes seconds. On CPU-only systems, it can take minutes per image. The choice of GPU depends on your existing hardware: NVIDIA is the most widely supported on Linux, while Apple Silicon is the obvious choice for Mac users. If you’re just experimenting, start with CPU — you can always add a GPU later.

7. Can I use other models besides Stable Diffusion?

Yes, Docker Model Runner supports multiple models available as DDUF packages on Docker Hub. While stable-diffusion is a good starting point, you can pull other variants or newer architectures as they become available. The command docker model pull <model-name> works for any model published in the ai/ namespace (e.g., ai/stable-diffusion-xl). Keep in mind that larger models require more RAM and GPU memory. Always check the model’s size and requirements before pulling. And because the API is OpenAI-compatible, Open WebUI will work with any image generation model that exposes the /v1/images/generations endpoint.

8. How do I verify everything is working correctly?

After pulling the model and launching Open WebUI, open your browser to the address displayed in the terminal (usually http://localhost:8080). In the chat interface, type a simple prompt like “a cat wearing a hat” and press enter. If the image appears within a reasonable time, everything is set up correctly. You can also run docker model list to see all locally available models, and docker model inspect stable-diffusion to confirm the model is intact. If you encounter errors, check Docker’s system requirements, ensure you have enough disk space for the model download (around 7 GB), and verify that your GPU drivers (if applicable) are up to date.

Tags: