Semantic Segmentation.

Definition

Semantic segmentation assigns a class label — water, grass, sky, building — to every single pixel of an image, and paints each class with its own flat colour. The result is a segmentation map: a picture of what is where, stripped of texture and light. It differs from instance segmentation, which tells apart individual objects (person 1 from person 2), and from panoptic segmentation, which does both. Class palettes are fixed by the dataset — ADE20K (150 classes), COCO-Stuff, Cityscapes — so the same blue always means water. It matters for design because the map is editable and reversible: a generative model can turn a hand-painted map back into a photograph, which is how NVIDIA's GauGAN worked and how ControlNet conditions diffusion models today.