Creative generative workflows built with Modular Diffusers ๐งจ for state-of-the-art open source models.
Diffusion pipelines are usually monolithic: one class per task, and a new task means a new pipeline. Modular
Diffusers breaks that into composable blocks โ encode, prepare, denoise, decode โ that you assemble into a
workflow, share components between them through a ComponentsManager so one loaded model serves several tasks,
and swap or subclass a single block instead of forking a pipeline. It is how a workflow that would be a graph of
custom nodes becomes a few lines of Python.