Flow-OPD

Flow-OPD: On-Policy Distillation for Flow Matching Models — Evaluated on SD-3.5-Medium, Flow-OPD achieves +18pt average improvement over vanilla GRPO.

Quick Start

import torch
from diffusers import StableDiffusion3Pipeline
from peft import PeftModel

model_id = "stabilityai/stable-diffusion-3.5-medium"
lora_ckpt_path = "CostaliyA/Flow-OPD"#dev ckpt
device = "cuda"

pipe = StableDiffusion3Pipeline.from_pretrained(model_id, torch_dtype=torch.float16)
pipe.transformer = PeftModel.from_pretrained(pipe.transformer, lora_ckpt_path)
pipe.transformer = pipe.transformer.merge_and_unload()
pipe = pipe.to(device)

prompt = "a photo of a black kite and a green bear"
image = pipe(prompt, height=512, width=512, num_inference_steps=40, guidance_scale=4.5, negative_prompt="").images[0]
image.save("flow_opd.png")

Results

Model	GenEval	OCR	DeQA	PickScore	Average
SD-3.5-M (base)	0.63	0.59	4.07	21.64	0.72
GRPO-Mix	0.73	0.83	4.33	21.84	0.82
Flow-OPD	0.92	0.94	4.35	23.08	0.90

Citation

@misc{fang2026flowopdonpolicydistillationflow,
      title={Flow-OPD: On-Policy Distillation for Flow Matching Models},
      author={Zhen Fang and Wenxuan Huang and Yu Zeng and Yiming Zhao and Shuang Chen and Kaituo Feng and Yunlong Lin and Lin Chen and Zehui Chen and Shaosheng Cao and Feng Zhao},
      year={2026},
      eprint={2605.08063},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.08063},
}