Flow-OPD: On-Policy Distillation Pushes GenEval +29 Points on Stable Diffusion 3.5
Flow-OPD is the first framework to integrate on-policy distillation into flow matching text-to-image models. A two-stage strategy — single-reward GRPO fine-tuning of specialized teacher models, then consolidation via dense trajectory-level vector field supervision with Manifold Anchor Regularization — achieves GenEval +29 points (63→92) and OCR accuracy +35 points (59→94) on Stable Diffusion 3.5 Medium, surpassing individual teacher models.
Why it matters
113 HF Daily upvotes; offers a principled solution to multi-objective RLHF alignment for diffusion models — a major open problem for production text-to-image systems attempting to satisfy competing objectives simultaneously.
Importance: 3/5
113 HF Daily upvotes; GenEval +29 on SD3.5 Medium via a principled multi-objective RLHF approach; first on-policy distillation framework for flow matching models.