Stable Diffusion gains Flax support in Hugging Face Diffusers for TPU inference

AI Impact Summary

HF Diffusers now supports Stable Diffusion in Flax, enabling TPU-optimized inference. The guide demonstrates loading CompVis/stable-diffusion-v1-4 with bf16 weights, replicating model parameters across 8 TPU devices, and generating 8 images in parallel, which can dramatically increase throughput for batch-style generation. Adoption requires upgrading to Diffusers >= 0.5.1, accepting the CreativeML OpenRAIL-M license, and provisioning a TPU-enabled environment (Colab/Kaggle/GCP) with proper HuggingFace authentication. Migration considerations: if you rely on PyTorch-based pipelines, you can target Flax for TPU, but you must adjust for Flax's stateless design and device-scoped RNG.