Local image generation models
This is the model index for local image generation. Each family page here answers the same set of questions so you can compare them without reading marketing copy: who makes it, what it is known for, which sizes exist and what hardware they fit on, which apps load it, and what the community has built on top of it.
Getting started with local image generation covers the setup this assumes, if you have not done it yet. The model choice matters less than having a working setup.
Families
- Stable Diffusion — the family the open image generation ecosystem grew up around. Multiple generations of architecture, the widest selection of community fine-tunes and adapters, and variants that run on modest hardware.
- Flux — a newer family known for prompt adherence and text rendering, with higher memory expectations and a growing adapter ecosystem of its own.
Families, fine-tunes and adapters
Three words get used loosely and they mean different things.
A family is a base architecture released by the people who trained it. Everything downstream inherits its shape, its text encoders and its memory footprint. When a page here says a model "fits", it is talking about the base weights plus encoders.
A fine-tune is that base retrained further on additional data, producing a new complete set of weights. Most community checkpoints you will find are fine-tunes: a photographic model, an illustration model, a model tuned for a particular subject. See fine-tuning. A fine-tune is a full model file, so it costs the same memory as the base and you load it instead of the base, not alongside it.
An adapter is a small file applied on top of a base model at load time. The most common kind is a LoRA, which stores a low-rank modification of the base weights rather than a whole new copy. Adapters are small, you can stack several, and you can turn them on and off per generation. That makes them the normal way to add a style, a character or a specific look without keeping a separate multi-gigabyte checkpoint for each one.
The practical consequence: adapters are tied to the family and often to the specific generation they were trained against. A LoRA built for one family will not load against another, and file names rarely tell you which. Check what a download says it targets before assuming it is broken.
Choosing between them
Decide on memory first. Look at your card on the GPU pages, then pick the family whose common variants fit with room to spare. After that, the ecosystem matters more than the base model, because whatever you generate day to day will probably be a fine-tune or a base plus adapters rather than the raw release.