Michael Limberger

Need me? Email mike@limberger.ca

In the room

This talk is about running Stable Diffusion on your own machine. AUTOMATIC1111 is the web UI most people use as the dashboard. The rest of the series covers install, models, and prompting basics.

That openness changed the room. People downloaded it, studied how it works, fine-tuned it, and shared what they made. A whole ecosystem popped up almost overnight.

The brain and the dashboard

These are not the same thing.

Stable Diffusion is the model itself: the brain. AUTOMATIC1111 (people say A1111) is a web interface that controls it: the dashboard. A1111 is a community project, not made by Stability AI. It is just the most popular way to drive the model.

Other dashboards exist. ComfyUI is node-based and more technical. Forge is an A1111 fork tuned for speed. InvokeAI is a polished UI with a different workflow.

Tonight we use A1111 because most documentation assumes it, most model pages give A1111 settings, the community and the extensions are the biggest, and the interface is beginner-friendly.

Training data, said plainly

Cloud services like DALL-E and Midjourney heavily filter their outputs. You type a prompt, you get a clothed person. That is by design.

Local Stable Diffusion models are a different story. Many community models are trained on uncensored datasets. Stability AI keeps plausible deniability. The models learned from everything.

Default outputs may not include clothing unless you ask for it. You will likely need a negative prompt (a list of things to stay away from) with tags like nude, nsfw if you want safe-for-work results consistently. This is not a bug. It is how the training data shaped the models.

None of this is a value judgment. Know what you are working with. If you are demoing at work or showing family, test your prompts first. The negative prompt part of this session covers that.

Three generations

SD 1.x landed in 2022 as the original release. It makes 512 by 512 images, runs on consumer GPUs, and was good for 2022, though it is showing its age now.

SD 2.x came late 2022. Better quality, but it broke compatibility with community tools. Nobody really adopted it.

SDXL is 2023, and this is what we use. Current generation. Native 1024 by 1024. Way better at composition, anatomy, and text. Models are bigger, about 6.5 GB versus about 2 GB, but the results speak for themselves.

Why the open weights matter

Because Stability released everything openly, the community built specialized versions. Want anime? Someone trained a model for that. Photorealistic portraits? Got it. Western comics? Multiple options.

The models in this session are community fine-tunes built on SDXL, each dialed in for a different look.

Where it lives

The A1111 repo is github.com/AUTOMATIC1111/stable-diffusion-webui. Once it is running, the WebUI sits at http://127.0.0.1:7860 on your own machine.