I’m not here to argue about the controversy surrounding it picture or video generation driven by generative AI models. Artists should be paid for their work. But if you’re going to use it LLM and their equivalentHere’s how to run the tools locally. It keeps your data private, lets you create images that commercial tools won’t let you, and you don’t have to pay a monthly subscription fee.
You don’t need a particularly powerful graphics card to do this. The tool I use works best Nvidia graphics cards With 8–12 GB of VRAM, but as with any AI, the more VRAM the better. You can use only CPU generation, but it’s slow. Like “go and brew a cup of tea slowly”, but “go to sleep and slowly see if you are ready in the morning”. But even with that limitation, it’s fun to learn the different settings that allow you to create an image, and that’s something you don’t get to experiment with much with a tab-limited subscription.
Want to stay up-to-date on the latest developments in artificial intelligence? The XDA AI Insider newsletter drops every week with deep dives, tool tips, and hands-on coverage you won’t find anywhere else on the site. subscribe by change your ballot.
What is SwarmUI and why would you use it?
Use your own hardware for imaging tasks
SwarmUI is a front-end for the popular ComfyUI rendering tool. But it’s more than that, because ComfyUI is complex and node-based, while SwarmUI simplifies things like the image generators you see online. This gives you the best of both worlds, as you can use a simple panel for quick images or get into the weeds with ComfyUI graphics.
Instead, what you get for installing SwarmUI is an image server built and installed for you. You get a Create tab that works like all simple image generators, so you don’t have to create a ComfyUI graphic every time you want a quick image. There is value in learning how to chart, but you don’t want to spend all that time if you don’t need to.
You also get power tools not available in ComfyUI, such as the Grid Generator, which lets you draw parameters side-by-side to see how changes to the recipe affect different aspects of the image being generated. Built-in image editor with built-in coloring and painting, image history browser with metadata search, and model browser to draw new models from Hugging Face and Civitai without leaving the UI.
It’s as easy as using Midjourney on the web, but it runs on your hardware. No calling to external servers, no telemetry, and no selling your data to anyone who will pay for it. Simply create an image that suits your conditions.
SwarmUI makes it easy to create a native image
The installation package does all the hard work for you
SwarmUI makes it easy to get up and running. There is a .bat file for Windowsor a script to pull for Linux users, and that’s it. For both, the script pulls all the code it needs, installs the server and necessary dependencies, and then opens the server in a browser tab to use.
The Windows script took less than five minutes to complete, and the SwarmUI server required a few minor details to complete. Things like the theme I want, what models to use and what backend (if you already have something installed).
Stable Diffusion XL 1.0 Base is about 6.5GB to boot and works fine, or Flux.1 Schnell is bigger and faster, but needs 12GB or more of VRAM to run. You also don’t need huge models unless you want to create a video. Here, the server pulls the models and configures them for you, and you can sit back for half an hour while they download in the background.
You can manipulate the Comfy UI if you want
It’s all there to deal with
The simplified Create tab is more like user-friendly online services. You choose which model to use, put a descriptive sentence or two in the information box, and add some things to avoid in the negative section. Then set a few other parameters, such as how many images to create per click, how many steps to take, and the aspect ratio you want.
What matters here is the CFG Size, which determines how strictly the model follows your instructions. Higher numbers are tighter, lower ones are more permissive, and anything between 5 and 8 will give you the best results (mostly). And Seed is usually set to -1 for a random number, but if you want to repeat the result, you can set it to any number you want.
But node-based Graphics created for ComfyUI is never unavailable and you can access as complex a workflow as you like. Since the main text-to-image flow is about seven nodes, you can get really detailed here. More involved images can be more and the content of each node is important to the overall image.
You get more control than you pay for with Midjourney or other tools, every option is yours. Sometimes it’s just fun to spend a few hours pushing nodes and adding values to see what happens, and it gives me a little more understanding of how models and the tools around them work.
Why pay $20 a month to create an image when you can do it locally
When GenAI was new, it was a decent idea to pay for imaging tools as native models struggled. That’s no longer the case, and self-hosted SwarmUI gets you comparable results without costing you a dime. Plus, you can run the prompts as many times as you want to get the results you want without any restrictions. It’s one of my go-to services now, and it helps the part of my brain that sees words, not pictures, to contextualize what it’s dreaming about.
