Stable Diffusion 3.5 Release

Let’s dive into the newly released Stable Diffusion 3.5 model. As one of the most powerful open-source AI image generation models available today, it’s designed to run on consumer hardware and is available under the Stability AI Community License Agreement. You can now download both Stable Diffusion 3.5 Large and Stable Diffusion 3.5 Large Turbo from Hugging Face, with the source code available on GitHub.

Why Upgrade to Version 3.5?

After releasing SD3 Medium in June, Stability AI acknowledged that it didn’t fully meet expectations. Instead of rushing patches, the team took time to develop this entirely new 3.5 version, responding to community feedback and creating a more robust solution for creators.

Available Models

Stable Diffusion 3.5 comes in three variants to suit different needs:

1. Stable Diffusion 3.5 Large

• 8.1 billion parameter model
• Optimized for professional use
• Supports 1 million pixel resolution
• The most powerful base model in the SD family

2. Stable Diffusion 3.5 Large Turbo

• Streamlined version of the Large model
• Generates high-quality images in just 4 steps
• Significantly faster than the Large version
• Maintains excellent prompt accuracy

3. Stable Diffusion 3.5 Medium

• 2.5 billion parameter model
• Features improved MMDiT-X architecture
• Optimized for consumer hardware
• Supports resolutions from 250K to 2M pixels

Technical Improvements and Optimizations

The new version brings significant improvements in several areas:

• Customization: Enhanced Query-Key normalization makes the model more adaptable for fine-tuning and development
• Hardware Friendly: Medium version runs with just 9.9GB VRAM (excluding text encoder)
• Balanced Performance: Achieves an optimal balance between image quality and prompt accuracy

Getting Started

Besides downloading model weights from Hugging Face for self-deployment, you can access these models through various platforms:

• Stability AI API
• Replicate
• DeepInfra
• ComfyUI