Flow Map Reward Guidance: Training-Free Generative Model Alignment
A new method called Flow Map Reward Guidance (FMRG) reformulates generative model guidance as a deterministic optimal control problem. FMRG uses the flow map—a concept from fast inference—to integrate and guide the flow in a single trajectory without training. At text-to-image scale, it matches or surpasses baselines in inverse problems, style transfer, and human preferences. The approach subsumes existing guidance methods at the coarsest level of a hierarchy of algorithms.
Key facts
- FMRG is a training-free, single-trajectory framework.
- It reformulates guidance as a deterministic optimal control problem.
- The flow map arises naturally in the optimal solution.
- FMRG matches or surpasses baselines at text-to-image scale.
- Applications include inverse problems, style transfer, and human preferences.
- Existing guidance methods require expensive multi-particle, many-step schemes.
- FMRG subsumes existing approaches at the coarsest level.
- The work is published on arXiv with ID 2604.27147.
Entities
Institutions
- arXiv