DiffusionOPD: A Unified Perspective of On-Policy Distillation in Diffusion Models

TL;DR

DiffusionOPD introduces a multi-task diffusion model training method using online policy distillation, effectively combining task-specific knowledge into a unified model with improved efficiency and performance.

Contribution

The paper presents a novel multi-task training paradigm for diffusion models based on online policy distillation, extending OPD to continuous Markov processes and demonstrating superior results.

Findings

DiffusionOPD outperforms multi-reward RL and cascade RL baselines in efficiency and performance.

Theoretical derivation of a closed-form KL objective unifies SDE and ODE refinements.

Empirical results achieve state-of-the-art benchmarks across multiple tasks.

Abstract

Reinforcement learning has emerged as a powerful tool for improving diffusion-based text-to-image models, but existing methods are largely limited to single-task optimization. Extending RL to multiple tasks is challenging: joint optimization suffers from cross-task interference and imbalance, while cascade RL is cumbersome and prone to catastrophic forgetting. We propose DiffusionOPD, a new multi-task training paradigm for diffusion models based on Online Policy Distillation (OPD). DiffusionOPD first trains task-specific teachers independently, then distills their capabilities into a unified student along the student own rollout trajectories. This decouples single-task exploration from multi-task integration and avoids the optimization burden of solving all tasks jointly from scratch. Theoretically, we lift the OPD framework from discrete tokens to continuous-state Markov processes, deriving a closed-form per-step KL objective that unifies both stochastic SDE and deterministic ODE refinement via mean-matching. We formally and empirically demonstrate that this analytic gradient provides lower variance and better generality compared to conventional PPO-style policy gradients. Extensive experiments show that DiffusionOPD consistently surpasses both multi-reward RL and cascade RL baselines in training efficiency and final performance, while achieving state-of-the-art results on all evaluated benchmarks.