Diffusers 0.37.0 adds Modular Diffusers, new image/video pipelines, and core library improvements

AI Impact Summary

Diffusers 0.37.0 introduces Modular Diffusers, enabling block-based assembly of diffusion pipelines that speeds experimentation and bespoke workflow creation alongside the existing DiffusionPipeline API. The release adds a slate of new models and pipelines across image, video, and audio (Z-Image Omni Base, Flux2 Klein, Qwen-Image-Layered, FIBO Edit, Cosmos Predict2.5, Cosmos-Transfer2.5, GLM-Image, RAE, LTX-2, Helios), expanding production-ready capabilities and editable content workflows. Core library improvements (caching, context-parallel backends, pipeline-level device_map, Transformers v5 compatibility) improve performance and deployment ergonomics, while production-focused features in FIBO Edit highlight deterministic editing. Teams should plan for hardware requirements (Flux2 Klein can run on 13GB VRAM; Helios/video workloads) and potential migration efforts to adopt modular blocks in existing pipelines.