Stable Diffusion: the family that founded local image generation
Stable Diffusion democratised image generation on your own machine and gave rise to today's common practices: checkpoints, LoRAs, ControlNet, image to image, inpainting, ComfyUI. What the family is, why its older generations are still in use, and where to read its sources.
Stable Diffusion is the family of models from Stability AI that democratised local image generation: its ecosystem gave rise to checkpoints, LoRAs, ControlNet, image to image, retouching by area and ComfyUI workflows. It is not a single model but several generations, with different architectures, capabilities and licences; the older ones remain in use thanks to the ecosystem built around them. At Axiris, Stable Diffusion is the family through which you concretely understand the operations of a diffusion pipeline, even when production runs on more recent engines.
The founding family of local generation
Stable Diffusion played a major role in democratising image generation run on your own machine. Its ecosystem gave rise to or popularised practices that are common today: checkpoints, LoRAs, ControlNet, image to image, retouching by area (inpainting), local interfaces, ComfyUI workflows, derived models.
A large part of the vocabulary still used for image models comes from there.
Several generations
Stable Diffusion is not a single model. Several generations and variants have been released by Stability AI, with different architectures, capabilities and licences: you read the licence of the model you use, not that of the family (AI model licences).
The older generations remain in use in some workflows, thanks to the considerable ecosystem built around them: a very large number of derived models, LoRAs and control tools that do not yet exist, or not to the same degree, for more recent families.
Stable Diffusion and LoRAs
The Stable Diffusion ecosystem did much to popularise model customisation through LoRAs: specialising a model around a style, a character, an aesthetic, a subject, without fully retraining it. What a LoRA is and when it is useful: LoRA: specialising a model. How to build one that holds up in production: LoRA workflow.
Stable Diffusion and ComfyUI
ComfyUI lets you use the various generations of Stable Diffusion in fully customisable workflows. Its node-based architecture makes the operations of a diffusion pipeline very concrete: loading the model, encoding the text, sampling, decoding, applying a LoRA, guiding with ControlNet. It is often on this family that you learn to read a graph: Nodes in ComfyUI.
Stable Diffusion at Axiris
Our production runs today on more recent engines; Models of the moment says which ones, for what and since when. Stable Diffusion remains the family through which you understand what a diffusion model does, and a serious option when a project depends on an ecosystem of LoRAs or controls that it alone possesses. The choice is made on the job, not on the age of the model: How to choose an AI image model?.
Official sources
Stability AI: image models · Stability AI · Official organisation on Hugging Face
Contact request: a first conversation, free and with no commitment. Every practice is different; it is by listening to you that we propose what suits you.