Stable Diffusion models for local use
Stable Diffusion is a family of open-weight image models from Stability AI, and it is the family listed first on the image models index for a reason: most of the tooling for local image generation was built against it. If an app, adapter format or workflow exists in the local image space, it almost certainly supports Stable Diffusion.
What the family is known for is reach rather than any single capability. Weights are downloadable, several architecture generations exist side by side, and the smaller variants run on hardware that cannot touch anything newer. That combination is why so much community work accumulated here.
Sizes and what they fit
The family spans several architecture generations, and they are not interchangeable in memory terms. Rather than treating it as one model, think of it as a ladder.
- Earlier generations are the lightweight end. Base weights plus text encoders fit in a few gigabytes of VRAM, which puts them within reach of older and mid-range cards, integrated graphics with enough shared memory, and laptops. These are the variants to reach for if you are working around a capacity ceiling.
- Later generations raised both the native resolution and the size of the text encoders. They want noticeably more memory, and they benefit from it: you get better output at higher resolutions without upscaling passes.
Capacity sets what you can attempt and memory bandwidth sets how fast each step runs, so two cards with the same amount of memory can feel very different. See the GPU pages for where specific cards land.
On a 24 GB card like the RTX 4090, every generation of this family fits with room left over for large batches, high resolutions and several adapters loaded at once. That headroom is what makes a card like this comfortable rather than merely sufficient.
On Apple silicon such as the M3 Max, the unified memory pool means capacity is rarely the binding constraint, and a well-specified machine can hold variants that would not fit on a discrete card of nominally similar class. Throughput per step is lower than a comparable desktop GPU, so generation takes longer while remaining entirely workable.
If you want to work the numbers for a specific file, see will this model fit in memory?.
Runtimes
Every app covered here loads this family, and it is usually the default they are configured around.
- ComfyUI exposes each stage of the pipeline as a node, which is the most direct way to combine multiple adapters, custom samplers and multi-pass workflows.
- SD WebUI grew up alongside this family. Its conventions for prompt weighting and adapter syntax became the informal standard that much of the community writes for.
- InvokeAI supports the family with a canvas-oriented workflow, suited to inpainting and iterating on a region of an existing image.
Checkpoint files are portable between them. Moving a model between apps is a matter of pointing the right directory at it, not converting anything.
Ecosystem
This is the family's real advantage. Around each architecture generation there is a large body of community work:
- Fine-tunes. Full checkpoints retrained for a particular look: photographic, illustrated, anime, architectural, product photography. See fine-tuning. You load one instead of the base.
- LoRA adapters. Small files that add a style, subject or concept on top of a base, stackable and toggleable per generation. Most style-specific work is distributed this way.
- Control and conditioning add-ons. Modules that steer composition from a reference image, a depth map or a pose, rather than from the prompt alone.
- Upscalers and refiners. Second-pass models used to raise resolution beyond what the base generates natively.
The one recurring trap is generation compatibility. An adapter trained against one architecture generation will not work with another, and file names often do not say which. Check the source page before assuming a download is faulty.