Stable Diffusion 3.5

Stable Diffusion 3.5 from Stability AI generates images from text prompts using a Multimodal Diffusion Transformer (MMDiT) architecture. The family ships in three open-weight variants: Large at 8 billion parameters, Large Turbo for faster distilled sampling, and Medium at 2.5 billion parameters, and the reference repository also runs SD3 Medium. Conditioning combines three text encoders, OpenAI CLIP-L/14, OpenCLIP bigG, and Google's T5-XXL, with a 16-channel VAE decoder, which improves text rendering, composition, and prompt adherence over SDXL. The inference code supports ControlNet guidance for blur, Canny edge, and depth conditioning, batch generation from prompt files, and a Skip Layer Guidance option for better coherency. The reference implementation code is MIT licensed while the weights fall under the Stability Community License, free for research and for smaller businesses. It is widely used by hobbyists on consumer GPUs with roughly 8 to 16 GB of VRAM and by teams integrating open image generation into products.