AI BriefingDiffusers Adds Support for Stable Diffusion 3
Key point
Hugging Face's Diffusers library now officially supports the Stable Diffusion 3 (SD3) model.
Details
Hugging Face's Diffusers library now officially supports Stability AI's latest model, Stable Diffusion 3 (SD3). This release includes the Stable Diffusion 3 Medium model with 2 billion parameters, immediately available via the Hugging Face Hub.
SD3 adopts a new MMDiT (Multimodal Diffusion Transformer) architecture. This embeds text and image data into a common dimension so they can interact, providing improved information flow compared to the existing cross-attention approach. The model utilizes three text encoders—CLIP L/14, OpenCLIP bigG/14, and T5-v1.1-XXL—along with a 16-channel AutoEncoder.
It also applies the Rectified Flow Matching training method, which connects data and noise distributions in a straight line to maximize sampling efficiency. For this, a new scheduler called FlowMatchEulerDiscreteScheduler has been introduced, and it also supports adjusting the shift parameter for high-resolution image generation.
Users can use SD3 by updating diffusers to the latest version, and fine-tuning scripts utilizing Dreambooth and LoRA are also provided.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.