Conditional Diffusion Models

Conditional diffusion models learn to generate samples from a distribution specified by inputs such as images, labels, text, measurements, or physical parameters. Common approaches inject conditions into U-Nets or diffusion transformers through cross-attention, adaptive normalization, ControlNet branches, or inpainting, while classifier-free guidance adjusts conditional influence during sampling. They are used for image and video editing, medical reconstruction, time-series forecasting, scientific simulation, and molecular or engineering design, with evaluation spanning fidelity, condition accuracy, uncertainty, and physical validity. A central challenge is balancing faithful control with diversity and realism while keeping sampling efficient and ensuring models generalize to unseen conditions or domains.