AI BriefingDiffusers Supports SD 3.5
Key point
Hugging Face's Diffusers library officially supports the Stable Diffusion 3.5 Large model.
Details
The Stable Diffusion 3.5 (SD3.5) model has been released on the Hugging Face Hub, and it is now available through the Diffusers library.
This release includes two checkpoints:
- 8B Large model: The base model for high-performance image generation.
- Timestep-distilled model: A model capable of fast generation with fewer inference steps of 4-8.
Key architectural changes are as follows:
- QK normalization: Applying a standard technique for training large-scale transformer models.
- Dual attention layers: Using dual attention layers instead of single attention in MMDiT blocks to improve performance.
Users can install the latest version of diffusers, obtain model access through Hugging Face login, and then perform inference using StableDiffusion3Pipeline. Additionally, inference through Quantization and LoRA training are also supported to improve memory efficiency.
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.