Stability AI released Stable Diffusion 3.5 Large and Large Turbo on Hugging Face. Large has 8.1 billion parameters. Large Turbo uses Distillation & Synthetic Data from Large to generate an image in four sampling steps.
Stable Diffusion
Also known as: SDXL · Stable Diffusion XL · SD 3.5
Plain English
Stable Diffusion is a family of models that turns a written prompt into an image. Its downloadable weights let people generate images on their own hardware.
Technical Overview
The models generate images in a compressed latent space and decode the result to pixels. The v1 and SDXL models use a text-conditioned U-Net for denoising. SDXL adds a second text encoder and can hand the last denoising steps to a separate refiner.
Chronological Entries (3)
Stability AI released SDXL 1.0. The base model can generate an image on its own. In the two-stage pipeline it passes a still-noisy latent representation to a refiner trained for the final denoising steps.
Stability AI announced the public release of the v1 weights and recommended the v1.4 checkpoint. Its training sometimes omitted the text prompt so the same network could make both predictions needed for classifier-free guidance.
Direct Relations (2)
- ImplementsDiffusion Models
- UsesLatent Space & VAE