Diffusers enables Stable Diffusion 3 Medium (2B) support with memory-optimized inference

AI Impact Summary

Stable Diffusion 3 Medium introduces a 2B-parameter SD3 with a multi-encoder architecture and new FlowMatch-based training, plus a FlowMatchEulerDiscreteScheduler for inference. The Diffusers integration adds memory-optimized paths: CPU offload, removal of the T5-XXL encoder during inference, and 8-bit quantization options, enabling SD3 on GPUs with lower VRAM. Access is gated on Hugging Face, requiring gate acceptance and login before use. This lowers hardware barriers for SD3 deployment but requires careful planning around gating, latency trade-offs from offloading, and accuracy implications when disabling encoders or using reduced-precision models.