Modular Diffusers: diffusion pipelines from composable blocks
Hugging Face Diffusers introduced ModularPipeline, which builds diffusion pipelines from reusable blocks (text encoding, image encoding, denoising, decoding) that can be inspected, swapped and run on their own.
What happened
Hugging Face introduced Modular Diffusers, a composable alternative to the classic DiffusionPipeline.
The technical details
From the post:
ModularPipeline.from_pretrained(...)defines the workflow without loading weights, andload_components(...)loads them, with dtype and quantisation settings- the pipeline is made of inspectable blocks. For FLUX.2 Klein 4B:
text_encoder→vae_encoder→denoise→decode - each block has its own inputs and outputs, can run independently, and can be added, removed or swapped
- custom blocks can be written and composed
- integration with Mellon, a node-based visual workflow tool for wiring blocks together
Why it matters
The blocks map one-to-one onto the Stable Diffusion anatomy from my July 2026 study post: text encoder, VAE, denoiser, decoder. Making them explicit turns “a pipeline” into something you can debug and remix.
Sources
This is a summary written from the linked public sources. Figures and numbers are as reported there.