I have tested more than a dozen AI image generators over the last year. Most of them want your email address before they show you a single sample. A few want a credit card. One wanted both, plus a phone number for “verification.” The tool I keep coming back to is the one that shows me a sample image first, lets me write a prompt, and never once asks me to log in.
That tool is Stable Diffusion running on my own hardware, and the value proposition is not subtle: zero accounts, zero per-image fees, zero rate limits, zero copyright strings attached to anyone else’s training set. The catch is the obvious one. You have to set it up. I am going to walk through what actually worked for me, on a Mac, on Linux, and on a Windows box, and what was a waste of time.
The two local stacks I keep coming back to
There are roughly three ways to run Stable Diffusion (an open-source text-to-image model) on your own machine today. I tried all three. The one I run daily is the simplest.
Automatic1111’s webui is the original local Stable Diffusion interface. It is a Python application that downloads its own dependencies, loads a model, and serves a browser UI on http://localhost:7860. You do not need to know what ComfyUI (a node-based workflow engine for Stable Diffusion) is. You do not need to know what SDXL (Stable Diffusion XL, the second-generation model architecture) means. You click a button, type a prompt, and you get an image in about eight seconds on a midrange GPU.
ComfyUI is the second option. It is a node-based workflow editor where you wire up “load checkpoint” to “CLIP text encode” to “KSampler” to “VAE decode” and so on. It is more powerful. It is also a UI that requires you to know what you are wiring. I do not use it for daily image generation. I use it when I want to reproduce someone else’s specific pipeline exactly.
The third option is one of the hosted “you provide the API key” services. I tried two of them. They were fine, and they re-introduced the login wall I was trying to avoid.
For this article I am sticking with Automatic1111. It is the tool I have actually used for 11 months straight.
What the install looks like in 2026
The install is one shell script on Linux and macOS. On Windows it is a .bat file that does the same thing. The script clones the webui repo, creates a Python virtual environment, downloads a PyTorch (a tensor computation library that runs the model on your GPU) build that matches your GPU driver, downloads the Stable Diffusion model weights (the trained parameters that turn text into pixels), and starts the server.
I timed the full thing on a 2021 MacBook Pro with an M1 Pro chip. Cold start to “first image” was 14 minutes, and 11 of those were the model download. The model file is 6.9 GB. The download is the slow part. After that, the next startup is about 40 seconds.
On Linux it was 9 minutes cold start, all on the same M1 Pro, using the same script, because Linux has faster filesystem and slightly less overhead on the dependency install. Windows was 17 minutes because the CUDA (Nvidia’s GPU compute library) install added steps.
The install is genuinely a one-time cost. After that you run one command and you have a local image generator.
How it actually behaves day to day
The model generates a 1024×1024 image in about 8 seconds on my M1 Pro. It uses about 7 GB of VRAM (video RAM on the GPU) and about 4 GB of system RAM while generating. When the generation is done, the image is in /tmp/stable-diffusion/output/, named with a timestamp.
I use it for four things, in this order of frequency:
- Blog header images. I type a 30-word prompt describing the topic, get 4 variations, pick the best. The whole loop takes 90 seconds.
- Concept art for a project I am scoping. I generate 30 to 50 variations of the same prompt to figure out what a “design language” might look like.
- Stickers for an open source release. I generate 20 versions, hand the best 3 to a designer.
- Silly gifts. I generated a portrait of my friend’s dog as a Roman senator. He loved it.
What I do not use it for: stock photography. The output looks “AI generated” enough that I do not want to use it in a paid context where the reader might think they are looking at a real photograph. The tool makes a different kind of art than stock photos. I treat it accordingly.
The models I have installed
The base model is SDXL 1.0, which is what ships with the installer. I also keep three fine-tuned models (community-trained versions of the base model, each specialized to a different visual style) on disk:
- RealVisXL. Photorealistic. The model I use when I want a result that does not look “AI generated.”
- DreamShaper. Stylized illustration. Good for blog headers.
- Proteus v0.4. The “do whatever the prompt says, even if it is weird” model. My favorite for one-off experiments.
Each model is 6 to 13 GB. Total disk cost is about 35 GB for the three plus the base.
I do not pay for any of them. All three are free downloads from Civitai (a community hub for sharing Stable Diffusion models). All three are licensed for non-commercial use. I am not a commercial user. If you are, read the license before you ship.
Trade-offs
Local Stable Diffusion is not free in time. The first install took 14 minutes, and the model downloads are 7 to 13 GB each. If you have a slow internet connection, the downloads are the actual cost. Time spent waiting on the model to download is time spent not generating images. Disk space is also a real cost. Three models plus the base is 35 GB.
The hardware floor is real too. You need a GPU with at least 6 GB of VRAM. The M1 Pro works. A modern Nvidia card with 8 GB works. Integrated graphics will not. I have tried to run the same model on a 2020 Intel MacBook Air with no discrete GPU. The result was a hard crash on the third generation.
In our case the math was obvious: zero recurring cost versus a $20 per month Midjourney subscription, after about two months. Your math will be different if you only need images once a quarter, or if you do not have a GPU-equipped machine at all.
If you generate images daily and you have a GPU, the local setup is a clear win. If you generate images twice a year, the install is more work than it is worth. The hosted tools exist for a reason.
Bottom line
The local Stable Diffusion install is the AI image generator that just works for me. No login, no per-image cost, no rate limit, no one’s training set being a copyright concern. The cost is one 14-minute install and the GPU you already own. If that trade-off works for you, the tool is the same one I have been running for 11 months without once hitting a paywall.
If I had to send a message back to the version of me that first installed it, I would say three things:
- Read the model license before you use a fine-tuned model in a commercial context. The community is good about permissive licenses, but “free to download” is not always “free to use in a paid product.”
- The first model you download is probably not the one you will keep. I tried nine. I kept three.
- Image generation is not a substitute for a designer. It is a substitute for “I need an image and I do not have time to brief a designer.” Use it accordingly.