Diffusers 0.36.0 adds Flux2/Z-Image pipelines, new video pipelines, kernels and TaylorSeer cache, plus LoRA training script

AI Impact Summary

Diffusers 0.36.0 expands the image and video generation surface with new pipelines (Flux2, Z-Image, QwenImage Edit Plus, Bria FIBO, Kandinsky Image Lite, ChronoEdit) and additional video pipelines (Sana-Video, Kandinsky 5.0 T2V Lite, HunyuanVideo-1.5, Wan Animate). It introduces kernel-backed attention backends (Flash Attention 3/2 and SAGE) and a TaylorSeer caching layer that can yield up to 3x speedups with minimal quality loss. A Flux.2 LoRA fine-tuning script and several refactors aim to streamline training and modularization; teams should validate compatibility with existing pipelines and update deployment/tests accordingly. Note: ChangeType indicates deprecation; verify if legacy components are slated for removal and plan migrations if needed.