Flow-Direct: Non-Parametric Guidance for Flow Models
Flow-Direct is a novel framework that requires no training for guiding pre-trained diffusion and flow models. It utilizes a continuous guidance field to enhance objectives specific to applications through external black-box reward functions. Unlike traditional approaches that overlook reward feedback after one application, Flow-Direct gathers all samples evaluated by rewards to create a non-parametric estimator of the guidance field. This estimator is theoretically based on the log-density ratio of the base distribution and the reward-weighted target distributions. By leveraging past feedback, this method significantly boosts feedback efficiency in guiding generation. Further details are available in a paper published on arXiv (2605.16348).
Key facts
- Flow-Direct is a training-free guidance framework.
- It uses a persistent guidance field for flow models.
- The guidance field is derived from the log-density ratio.
- It employs a non-parametric estimator from accumulated samples.
- Existing methods discard reward feedback after one use.
- Flow-Direct reuses historical reward feedback.
- The paper is on arXiv with ID 2605.16348.
- It targets pre-trained diffusion and flow models.
Entities
Institutions
- arXiv