Generative AI models learn to segment objects they've never seen before

Researchers found that AI models trained to generate images can be repurposed to identify and separate individual objects in photos, even when shown object types they never trained on. This works because the generative process forces the model to learn where objects begin and end, a skill that transfers to completely new categories.

Computer vision has relied on two different approaches: generative models that learn to create images, and discriminative models that learn to classify them. This paper shows the two aren't separate — the generative skill is more general. Models trained to generate furniture and cars can suddenly segment animals, plants, and other objects they never saw, approaching the performance of heavily supervised systems. The implication is simpler: if you want a model that works on novel objects, train it to generate, not classify.