Stable Diffusion Explains Latent Diffusion Image Generation

This Nov 2022 updated explainer breaks down how Stable Diffusion generates images from text, describing its CLIP text encoder, UNet-based latent diffusion process, and autoencoder decoder. It notes operational details—77×768 token embeddings, diffusion steps commonly set to 50–100, latent shape (4,64,64) decoded to (3,512,512)—and explains training with noisy-image denoising on datasets such as LAION Aesthetics, highlighting latent-space speed gains.