Stable Diffusion is categorically different from everything else: it is not a service but a model you download and run yourself.
That single difference creates both its biggest advantage and its biggest drawback. Once installed, everything is free — five images a day or five hundred costs the same. But setup is not easy and it wants a strong GPU.
This guide covers what to know before installing and, honestly, who this route suits.
First: is this route right for you?
If you generate occasionally, Stable Diffusion is the wrong tool. The setup and learning time costs more than several months of Midjourney.
Three situations have no substitute. First, showing the same product or person across scenes with full consistency — you teach the model that subject with a LoRA. Second, preserving a specific pose, depth or spatial geometry — ControlNet is the only method that does it. Third, images that must never leave your machine.
If none of those applies and you simply want beautiful images, Midjourney gives you better results with less effort.
Hardware: the realistic expectation
A GPU with 8 GB of VRAM runs basic generation. 12 GB or more makes larger models and higher resolutions comfortable.
It works on Apple Silicon Macs but is noticeably slower than an NVIDIA card in the same class — minutes rather than seconds per image.
Without a GPU the route is not entirely closed: services like Replicate run the same models in the cloud for a per-generation fee. That loses the free advantage but keeps the control advantage.
Two interfaces: which to choose
AUTOMATIC1111 is better known and offers a classic form interface: fill in the fields, press the button. Easier for a beginner to grasp.
ComfyUI is node-based. It is markedly harder to learn but makes complex workflows repeatable, and it has become the professional standard.
Practical advice: do your first install with AUTOMATIC1111 and learn the core concepts. When you find yourself repeating the same workflow, it is time to move to ComfyUI.
The real value: ControlNet and LoRA
What separates Stable Diffusion is not image quality but control, and these two extensions are what justify the steep learning curve.
ControlNet binds generation to a reference image's skeleton: a person's pose, a room's depth map, a drawing's edges. If you need the same product from exactly the same angle against ten backgrounds, this is the only method that does it.
LoRA is the cheap way to teach the model something new. Fifteen or twenty images teach it your face, your product or your brand's illustration style, and you invoke that subject in any prompt afterwards.
The variety of your training images determines the result: if they are all from one angle, the model learns the angle rather than the subject.
Frequently asked questions
Is Stable Diffusion really completely free?
Installed on your own machine, yes: the weights download free and you pay nothing per generation. The cost is hardware and electricity. Through cloud services you pay per generation.
What GPU do I need?
8 GB of VRAM minimum for basic generation; 12 GB or more makes larger models and higher resolutions comfortable. It runs on Apple Silicon but noticeably slower than an equivalent NVIDIA card.
Which model version can I use commercially?
Licences vary by version, and this is the most frequently missed detail. Read the licence of the exact model version you run before commercial work. If you use it through a platform, ask which version it runs — many do not state it.