Stable Diffusion is a pipeline of three neural networks, not a single monolithic model

Stable Diffusion is a modular system of three distinct neural-network components that run in sequence. ClipText encodes the text prompt into 77 token-embedding vectors of 768 dimensions each. A UNet plus a scheduling algorithm iteratively processes a noise tensor in latent (compressed) space to produce an information array matching the prompt. An Autoencoder Decoder then converts that latent array (4×64×64) into a pixel image (3×512×512). Understanding SD as three cooperating models rather than one black box matters because each component can be swapped independently — a better text encoder or a different scheduler changes a different part of the pipeline. It also explains why prompt wording, step count, and model variants each affect a different stage.

Examples

Dimensions: ClipText output = 77×768. UNet works on a 4×64×64 latent tensor conditioned on the text embeddings. Autoencoder decoder output = 3×512×512 pixel image.

Assessment

Draw a block diagram of Stable Diffusion with three components, labelling each component’s input and output shapes. Explain which component you would change to improve text understanding versus final image quality.