Conditional Diffusion with Less Explicit Guidance via Model Predictive Control

TL;DR

This paper introduces a model predictive control approach to improve conditional diffusion sampling with limited explicit guidance, enhancing generative quality by simulating unconditional diffusion and backpropagating guidance feedback.

Contribution

It proposes a novel MPC-like method to approximate guidance in conditional diffusion models, reducing the need for extensive explicit guidance.

Findings

MPC guides closely match real guidance with high cosine similarity.

Adding MPC steps improves quality with limited guidance.

Effective guidance approximation over large simulation distances.

Abstract

How much explicit guidance is necessary for conditional diffusion? We consider the problem of conditional sampling using an unconditional diffusion model and limited explicit guidance (e.g., a noised classifier, or a conditional diffusion model) that is restricted to a small number of time steps. We explore a model predictive control (MPC)-like approach to approximate guidance by simulating unconditional diffusion forward, and backpropagating explicit guidance feedback. MPC-approximated guides have high cosine similarity to real guides, even over large simulation distances. Adding MPC steps improves generative quality when explicit guidance is limited to five time steps.