What parts does Stable Diffusion (SD) consist of?

Stable Diffusion consists of several key components:

1. Conditional Autoencoder — encodes the input image into a compact latent space and decodes it back, allowing work with smaller data sizes.

2. Diffusion Model — learns to gradually add noise to the latent representation and then restore the original image, enabling the generation of new images from noise.

3. Conditioning mechanism — usually a text encoder (e.g., CLIP) that transforms a textual description into a vector influencing the generation process to create images matching the description.

4. Diffusion and reverse diffusion sampling process — an algorithm that iteratively transforms random noise into a meaningful image using the trained model.

Here's an approximate schematic of the process:

Text description -> Text encoder -> Condition -> Diffusion model -> Latent space -> Decoder -> Image