Diffusion Guidance

Diffusion guidance studies how to steer a pretrained denoising or flow model during sampling toward a condition or objective, enabling controllable generation and inverse-problem solutions without rebuilding the generator. Common approaches include classifier-free and contrastive score guidance, gradients from rewards, energies, or measurement residuals, and particle or optimization methods; applications range from text-to-image alignment and image restoration to robot motion, molecular design, and safe planning. Recent work also uses adaptive schedules, uncertainty- or geometry-aware corrections, and internal model features to reduce extra evaluations and handle nondifferentiable or spatially localized constraints. A central tension is balancing condition satisfaction against fidelity, diversity, and staying on the learned data manifold, while controlling sensitivity to guidance weights, approximate gradients, and sampling cost.