Diffusion Model Guidance

Diffusion model guidance studies how to steer reverse denoising so generated samples satisfy goals such as matching text or images, honoring observations, meeting physical or safety constraints, and avoiding unwanted content. Common approaches include classifier-free and classifier guidance, gradients from reward or measurement models, attention and representation edits, and particle-based methods such as sequential Monte Carlo and Doob h-transforms. These techniques are used across text-to-image and video generation, image restoration, molecular design, and robot planning, with evaluation ranging from compositional-alignment benchmarks to inverse-problem fidelity and task success. A central tension is that stronger or more elaborate guidance can improve control but also damage realism, diversity, or fidelity to the model’s learned distribution, while adding computation; no alternative to standard classifier-free guidance is consistently best across settings.