Stable Diffusion 3.5 Large Turbo

Stable Diffusion 3.5 Large Turbo is a distilled, high-speed text-to-image model released by Stability AI. It is an 8.1 billion parameter model built on the Multimodal Diffusion Transformer (MMDiT) architecture, specifically optimized for rapid inference. By leveraging Adversarial Diffusion Distillation (ADD), the model can generate high-quality images in as few as 4 steps, significantly reducing the computational time required compared to the non-distilled Large variant.

The model's architecture utilizes three fixed, pretrained text encoders—OpenCLIP-ViT/G, CLIP-ViT/L, and T5-XXL—integrated with Query-Key (QK) normalization to enhance training stability. These components allow the model to maintain high levels of prompt adherence and spatial accuracy. It is particularly effective at rendering typography and managing complex, multi-element prompts that describe specific layouts or intricate textures.

For optimal results, users are encouraged to generate images at approximately 1-megapixel resolution (such as 1024x1024). Because the model is distilled, it typically performs best with a Classifier-Free Guidance (CFG) scale of 1.0; using higher CFG values may lead to visual artifacts. Despite its focus on speed, it remains highly customizable and supports the development of fine-tuned adapters and LoRAs for specialized creative tasks.