Introducing Modular Diffusers - Composable Building Blocks for Diffusion Pipelines

Modular Diffusers introduces a composable block-based approach to building diffusion pipelines, replacing monolithic pipeline classes with reusable, swappable components.

• •Pipelines are composed of self-contained blocks (text encoding, VAE encoding, denoising, decoding) that can be inspected, popped, and run independently via init_pipeline()
• •ModularPipeline.from_pretrained separates workflow definition from model weight loading, with load_components() handling dtype and quantization
• •Custom blocks are Python classes defining expected_components, inputs, intermediate_outputs, and a __call__ method for computation logic
• •ComponentsManager handles memory across multiple pipelines by automatically offloading models to CPU when not in use
• •