Conditional Generation

Conditional generation studies how to produce outputs that satisfy a given condition, such as a class label, molecular property, text prompt, trajectory, or physiological signal. Common approaches include conditional GANs, diffusion and score models, rectified flows, and autoregressive models, with guidance or preference optimization added for finer control. Applications range from images, video, and dialogue to molecules, medical signals, and scientific simulations, and evaluation tests both output quality and whether requested attributes are met. A central challenge is balancing condition fidelity with realism and diversity, especially for rare or unseen combinations and across domains.