Syn-OPD: Conflict-Aware On-Policy Distillation for Generative Video Super-Resolution

Generative video super-resolution (VSR) has recently benefited from powerful video generative priors, enabling realistic detail synthesis beyond conventional reconstruction-based methods. However, a fundamental fidelity–generation dilemma remains: improving perceptual realism often comes at the expense of faithful reconstruction, while optimizing fidelity tends to suppress generative details. On-policy distillation (OPD) provides dense trajectory-level supervision that reduces unguided exploration, while multi-teacher variants enable capability integration across specialized experts. However, when fidelity-oriented and generation-oriented experts supervise the same restoration state, their optimization directions may conflict. This raises two coupled questions: **which expert should dominate at each noise level, and how should conflicting directions be reconciled?** To address these challenges, we propose **Syn-OPD**, a **conflict-aware multi-teacher OPD framework** that consolidates fidelity and generative capabilities into a single VSR model. We first specialize two complementary teachers from the same initialization: a **Fidelity Expert** for structure-preserving reconstruction and a **Generative Expert** for realistic detail synthesis. During distillation, hybrid SDE–ODE rollouts provide student-visited states for dense same-state supervision. Syn-OPD then resolves expert competition progressively: timestep-aware gating determines which expert should dominate, while asymmetric gradient decoupling preserves the preferred direction and removes only the conflicting auxiliary component. Extensive experiments on six benchmarks demonstrate consistent gains in both perceptual quality and reconstruction fidelity, with our 1.3B model even outperforming larger VSR models on perceptual metrics.