# U-Net **U-Net** is a convolutional neural network architecture built around two mirrored halves: a contracting *encoder* path that compresses an image into a compact feature representation, and an expansive *decoder* path that reconstructs spatial detail back out; with **skip connections** directly wiring each encoder stage to its matching decoder stage so that fine-grained spatial information isn't lost during the squeeze. Originally designed for biomedical image segmentation (Ronneberger, Fischer & Brox, 2015), it became the backbone of modern diffusion models — [[stable_diffusion|Stable Diffusion]], [[dalle|DALL-E 2]], and [[midjourney|Midjourney]] all use a U-Net to iteratively predict and remove noise at each denoising step. ^overview > [!example] > Without skip connections, the decoder has to reconstruct *where* things are using only the bottleneck: resulting in blurry, spatially imprecise outputs. With skip connections, the decoder receives high-resolution feature maps directly from the encoder, letting it answer both *what is here* and *exactly where* in a single forward pass.