Decoder. plain-English AI glossary

U-Net

● Core

A U-shaped network that shrinks an image down, then grows it back — with shortcuts across the U so detail survives.

Think of it like

Like sketching the big shapes first, then zooming back in with your original reference photo taped beside you so you don’t lose the fine lines.

Example

The denoising step inside most image diffusion models is a U-Net predicting what noise to remove.

How it actually works

The downsampling half captures context (what’s in the picture), the upsampling half restores resolution (where exactly), and skip connections pass fine-grained detail directly across the U so it isn’t lost in the squeeze. Born in medical image segmentation, it became the default backbone for diffusion because it’s good at pixel-precise prediction.

For product teams

The engine room of most image generators, even if users never hear the name.

For engineers

An encoder–decoder CNN with symmetric skip connections between matching resolution levels.

Diffusion Model — The generative process it usually powers.
Skip Connection — The shortcut trick it relies on.

Read anything AI without the jargon

Look up any term in plain English, or save terms as you read with the free Chrome extension.

Open Decoder Add to Chrome