Flow-OPD: On-Policy Distillation for Flow Matching Models

2026-05-08

AI Summary

This work addresses the challenges of sparse rewards and gradient interference in multi-task aligned text-to-image generation, which often lead to conflicting objectives and reward hacking. To overcome these issues, the authors propose the Flow-OPD framework, which first fine-tunes domain-specific teacher models via single-task GRPO and then distills heterogeneous knowledge into a unified student model through flow matching with cold-start initialization, policy sampling, task-routing annotations, and trajectory-level supervision. Notably, this study introduces on-policy distillation into flow matching for the first time and proposes Manifold Anchor Regularization to preserve generation quality. Evaluated on Stable Diffusion 3.5 Medium, the method boosts the GenEval score from 63 to 92 and OCR accuracy from 59 to 94, outperforming the original GRPO by approximately 10 points while maintaining high image fidelity and alignment with human preferences.

Abstract

Existing Flow Matching (FM) text-to-image models suffer from two critical bottlenecks under multi-task alignment: the reward sparsity induced by scalar-valued rewards, and the gradient interference arising from jointly optimizing heterogeneous objectives, which together give rise to a 'seesaw effect' of competing metrics and pervasive reward hacking. Inspired by the success of On-Policy Distillation (OPD) in the large language model community, we propose Flow-OPD, the first unified post-training framework that integrates on-policy distillation into Flow Matching models. Flow-OPD adopts a two-stage alignment strategy: it first cultivates domain-specialized teacher models via single-reward GRPO fine-tuning, allowing each expert to reach its performance ceiling in isolation; it then establishes a robust initial policy through a Flow-based Cold-Start scheme and seamlessly consolidates heterogeneous expertise into a single student via a three-step orchestration of on-policy sampling, task-routing labeling, and dense trajectory-level supervision. We further introduce Manifold Anchor Regularization (MAR), which leverages a task-agnostic teacher to provide full-data supervision that anchors generation to a high-quality manifold, effectively mitigating the aesthetic degradation commonly observed in purely RL-driven alignment. Built upon Stable Diffusion 3.5 Medium, Flow-OPD raises the GenEval score from 63 to 92 and the OCR accuracy from 59 to 94, yielding an overall improvement of roughly 10 points over vanilla GRPO, while preserving image fidelity and human-preference alignment and exhibiting an emergent 'teacher-surpassing' effect. These results establish Flow-OPD as a scalable alignment paradigm for building generalist text-to-image models.

Problem

• Flow Matching
• multi-task alignment
• reward sparsity
• gradient interference
• reward hacking

Innovation

• On-Policy Distillation
• Flow Matching
• Manifold Anchor Regularization
• Multi-task Alignment
• Trajectory-level Supervision