We propose a general approach for modifying image-based models to improve robustness and temporal consistency on video.

We build a nearly-perfect predictive model for memorization in diffusion models using theory and controlled experiments.