Introducing Modular Diffusers - Composable Building Blocks for Diffusion Pipelines

Hugging Face’s Modular Diffusers introduces a new paradigm for building diffusion pipelines by composing reusable blocks rather than writing entire pipelines from scratch. The framework extends the existing DiffusionPipeline class with a flexible, composable alternative that allows users to mix and match blocks such as text encoding, image encoding, denoising, and decoding. Each block is self‑contained with its own inputs and outputs, can be run independently, and can be added, removed, or swapped freely; the system dynamically recomposes the pipeline to accommodate changes. Users can create custom blocks as Python classes defining components, inputs, outputs, and computation logic, and publish them to the Hub for others to load with trust_remote_code=True. Modular Diffusers also introduces modular repositories that reference components from original model repos, enabling quantized models and custom pipeline blocks to coexist in a single repository. The framework integrates with Mellon, a node‑based visual workflow interface that automatically generates node UIs from block definitions, allowing users to wire blocks together visually. Community members have begun publishing complete modular pipelines on the Hub, and existing Diffusers pipelines can be converted to modular form using ModularPipeline.from_pretrained. The release aims to provide the composability and flexibility requested by the community while maintaining the features that make Diffusers powerful, and invites user feedback to shape future development.