Adding reward curve of Flow-GRPO-Fast vs. Flow-GRPO. In Pickscore reward, Flow-GRPO-Fast is comparable to Flow-GRPO with only 2 steps training.
2025-08-04
Adding support for FLUX.1-Kontext-dev. For the counting task, we use Geneval reward to detect object counts and CLIP feature similarity to ensure consistency between the original and edited images. This implementation offers a runnable pipeline, but the training set contains only 800 samples. Making Flow-GRPO truly effective for editing tasks still requires further exploration by the community.
2025-07-31
Adding Flow-GRPO-Fast.
2025-07-28
Adding support for FLUX.1-dev.
Adding support for CLIPScore as reward model.
Introducing config.sample.same_latent to control whether the same noise is reused for identical prompts, addressing Issue #7.
To improve training efficiency, we provide a better set of parameters for Flow-GRPO.
We found the following adjustments significantly accelerate training:
No CFG during training or testing — the RL process effectively performs CFG distillation.
Use the window mechanism from Flow-GRPO-Fast or MixGRPO — only train on partial steps.
Adopt Coefficients-Preserving Sampling (CPS) — CPS provides a notable improvement on GenEval, and produces higher-quality samples. A typical setting is noise_level = 0.8, which works well without tuning for different models or step counts.
The figure below shows the test-set performance curves using GenEval and PickScore as rewards, where both training and evaluation are performed without CFG. The experiments are configured with geneval_sd3_fast_nocfg and pickscore_sd3_fast_nocfg, using scripts from scripts/multi_node/sd3_fast.
Flow-GRPO-Fast Illustration
Flow-GRPO-Fast Illustration
🛡️ Over-optimization (GRPO-Guard) 🔥🔥
To mitigates implicit over-optimization in flow matching, our team propose GRPO-Guard ( 🔥Project Page).
We first observe that the importance ratio exhibits an inherent bias:
Its mean is consistently below 1 and becomes significantly pronounced at low-noise steps (e.g., step 8 in SD3.5-M).
The variance varies notably across different steps.
Ideally, the importance ratio distribution should have a mean of 1 and stable variance. The clipping operation truncates overly confident positive or negative samples outside the region [1−ϵ,1+ϵ], ensuring stable gradient updates. However, the observed bias in the importance ratio disrupts this mechanism—gradients of positive samples are no longer properly constrained, leading the policy model into over-optimization. As a result, the proxy score continues to rise while the gold score declines, causing a severe degradation in image quality.
The biased ratio distributions are summarized in the table below.
FlowGRPO
GRPO-Guard
flow_grpo ratio
grpo_guard ratio
The clipping mechanism is imbalanced, failing to constrain overconfident positive samples.
The clipping mechanism is imbalanced, failing to constrain overconfident positive samples.
To address this issue, GRPO-Guard introduces two mechanisms that effectively alleviate over-optimization:
RatioNorm: Corrects the distributional bias of importance ratios and unifies their statistics across denoising steps.
Gradient Reweight: Further reweights the gradients of different denoising steps based on RatioNorm, balancing their contributions and preventing excessive optimization under specific noise levels.
The following figure compares over-optimization between GRPO-Guard and FlowGRPO on text rendering tasks. GRPO-Guard maintains the same rising trend in proxy scores as FlowGRPO while preventing rapid declines in gold scores, thus preserving high image quality and diversity.
GRPO-Guard Illustration
Start Training
After downloading the base model and setting up the reward model, run the following script to start training the GRPO-Guard for the SD3.5-M text rendering task.
bash
1# Master node2bash scripts/multi_node/sd3_grpo_guard.sh 03# Other nodes4bash scripts/multi_node/sd3_grpo_guard.sh 1
Flow-GRPO-Fast
We propose Flow-GRPO-Fast, an accelerated variant of Flow-GRPO that requires training on only one or two denoising step per trajectory. For each prompt, we first generate a deterministic trajectory using ODE sampling. At a randomly chosen intermediate step, we inject noise and switch to SDE sampling to generate a group. The rest of the process continues with ODE sampling. This confines stochasticity to one or two steps, allowing training to focus solely on that steps. This few-step training idea was primarily proposed by Ziyang Yuan during our discussions in early June.
Each trajectory is trained only once or twice, significantly reducing the training cost.
Sampling before branching requires only a single prompt without group expansion, further speeding up data collection.
Experiments on PickScore show that Flow-GRPO-Fast matches the reward performance of Flow-GRPO while offering faster training speed. The x-axis in the figure represents training epochs. Flow-GRPO-Fast with 2 training steps per iteration performs better than Flow-GRPO, while Flow-GRPO-Fast with only 1 training step per iteration performs slightly worse than Flow-GRPO. In both cases, compared to Flow-GRPO’s 10 training steps per iteration, the training process is significantly faster.
Flow-GRPO-Fast Illustration
Please use scripts in scripts/multi_node/sd3_fast to run these experiments.
To avoid redundant downloads and potential storage waste during multi-GPU training, please pre-download the required models in advance.
Models
SD3.5: stabilityai/stable-diffusion-3.5-medium
Flux: black-forest-labs/FLUX.1-dev
Reward Models
PickScore:
laion/CLIP-ViT-H-14-laion2B-s32B-b79K
yuvalkirstain/PickScore_v1
CLIPScore: openai/clip-vit-large-patch14
Aesthetic Score: openai/clip-vit-large-patch14
3. Reward Preparation
The steps above only install the current repository. Since each reward model may rely on different versions, combining them in one Conda environment can cause version conflicts. To avoid this, we adopt a remote server setup inspired by ddpo-pytorch. You only need to install the specific reward model you plan to use.
GenEval
Please create a new Conda virtual environment and install the corresponding dependencies according to the instructions in reward-server.
PickScore requires no additional installation. Note that the original pickscore dataset corresponds to dataset/pickscore in this repository, containing some NSFW prompts. We strongly recommend using pickapic_v1_no_images_training_sfw, the SFW version of the Pick-a-Pic dataset, which corresponds to dataset/pickscore_sfw in this repository.
DeQA
Please create a new Conda virtual environment and install the corresponding dependencies according to the instructions in reward-server.
UnifiedReward
Since sglang may conflict with other environments, we recommend creating a new conda environment.
For Flow-GRPO-Fast, please use scripts/multi_node/flux_fast.sh. See the W&B logs for Geneval (with geneval_flux_fast in the config) and PickScore (with pickscore_flux_fast in the config).
Multi-node training for FLUX.1-Kontext-dev
Please first download generated_images.zip and extract it into the counting_edit directory. You can also use the scripts in the counting_edit directory to generate the data yourself.
Please install diffusers from the main branch to support FLUX.1-Kontext-dev:
After upgrading Diffusers, some packages such as PEFT may also need to be upgraded. If you encounter any errors, please upgrade them according to the error messages.
Then, run the scripts:
In the implementation of Qwen-Image, we have unified Flow-GRPO and Flow-GRPO-Fast. You can control the size of the SDE window with config.sample.sde_window_size, and adjust the position of the window with config.sample.sde_window_range.
Please install diffusers from the main branch to support Qwen-Image:
Using the provided configuration, the resulting reward curve of Qwen-Image on the test set is shown below.
Flow-GRPO-Fast Illustration
---
Multi-node training for Qwen-Image-Edit:
Same as Flux Kontext, please first download generated_images.zip and extract it into the counting_edit directory. You can also use the scripts in the counting_edit directory to generate the data yourself.
Please install diffusers from the main branch to support Qwen-Image-Edit:
Using the provided configuration, the resulting reward(PickScore) curve of Bagel on the test set is shown below (with 32 GPU).
Flow-GRPO-Fast Illustration
[Note]: About resource requirements & OOM
The default training script adopts full-parameter mode, whcih requires at least 8 × 80GB GPUs. If you encounter OOM issues, you can switch to LoRA training with the config provided in config/grpo.py:pickscore_bagel_lora.
Please update the entry Python script and config file names in the scripts/multi_node bash file.
FAQ
Please use fp16 for training whenever possible, as it provides higher precision than bf16, resulting in smaller log-probability errors between data collection and training. For Flux and Wan, becauase fp16 inference cannot produce valid images or videos, you will have to use bf16 for training. Note that log-probability errors tend to be smaller at high-noise steps and larger at low-noise steps. Training only on high-noise steps yields better results in this case. Thanks to Jing Wang for these observations.
When using Flow-GRPO-Fast, set a relatively small clip_range, otherwise training may crash.
When implementing a new model, please check whether using different batch sizes leads to slight differences in the output. SD3 has this issue, which is why I ensure that the batch size for training is the same as that used for data collection.
How to Support Other Models
To integrate a new model into this framework, please follow the steps below:
1. Add the following files adapted for your model:
flow_grpo/diffusers_patch/sd3_pipeline_with_logprob.py:
This file is adapted from pipeline_stable_diffusion_3.py. You can refer to diffusers for your model.
flow_grpo/diffusers_patch/sd3_sde_with_logprob.py:
This file handles SDE sampling. In most cases, you don't need to modify it. However, if your definitions of dt or velocity differ in sign or convention, please adjust accordingly.
2. Verify SDE sampling:
Set noise_level = 0 in sde_demo.py to check whether the generated images look normal. This helps verify that your SDE implementation is correct.
3. Ensure on-policy consistency:
Set config.sample.num_batches_per_epoch = 1 and config.train.gradient_accumulation_steps = 1 to enforce a purely on-policy setup, where the model collecting samples is identical to the one being trained.
Under this setting, the ratio should remain exactly 1. If it's not, please check whether the sampling and training code paths differ—for example, through use of torch.compile or other model wrappers—and make sure both share the same logic.
4. Tune reward behavior:
Start with config.train.beta = 0 to observe if the reward increases during training. You may also need to adjust the noise level here based on your model. Other hyperparameters are generally model-agnostic and can be kept as default.
🏁 Multi Reward Training
For multi-reward settings, you can pass in a dictionary where each key is a reward name and the corresponding value is its weight.
For example:
python
1{2"pickscore":0.5,3"ocr":0.2,4"aesthetic":0.35}
This means the final reward is a weighted sum of the individual rewards.
The following reward models are currently supported:
Geneval evaluates T2I models on complex compositional prompts.
OCR provides an OCR-based reward.
PickScore is a general-purpose T2I reward model trained on human preferences.
DeQA is a multimodal LLM-based image quality assessment model that measures the impact of distortions and texture damage on perceived quality.
ImageReward is a general-purpose T2I reward model capturing text-image alignment, visual fidelity, and safety.
QwenVL is an experimental reward model using prompt engineering.
Aesthetic is a CLIP-based linear regressor predicting image aesthetic scores.
JPEG_Compressibility measures image size as a proxy for quality.
UnifiedReward is a state-of-the-art reward model for multimodal understanding and generation, topping the human preference leaderboard.
✨ Important Hyperparameters
You can adjust the parameters in config/grpo.py to tune different hyperparameters. An empirical finding is that config.sample.train_batch_size * num_gpu / config.sample.num_image_per_prompt * config.sample.num_batches_per_epoch = 48, i.e., group_number=48, group_size=24.
Additionally, setting config.train.gradient_accumulation_steps = config.sample.num_batches_per_epoch // 2.
🤗 Acknowledgement
This repo is based on ddpo-pytorch and diffusers. We thank the authors for their valuable contributions to the AIGC community. Special thanks to Kevin Black for the excellent ddpo-pytorch repo.
⭐Citation
If you find Flow-GRPO useful for your research or projects, we would greatly appreciate it if you could cite the following paper:
@article{liu2025flow,
title={Flow-grpo: Training flow matching models via online rl},
author={Liu, Jie and Liu, Gongye and Liang, Jiajun and Li, Yangguang and Liu, Jiaheng and Wang, Xintao and Wan, Pengfei and Zhang, Di and Ouyang, Wanli},
journal={arXiv preprint arXiv:2505.05470},
year={2025}
}
If you find GRPO-Guard useful for your research or projects, we would greatly appreciate it if you could cite the following paper:
@misc{wang2025grpoguardmitigatingimplicitoveroptimization,
title={GRPO-Guard: Mitigating Implicit Over-Optimization in Flow Matching via Regulated Clipping},
author={Jing Wang and Jiajun Liang and Jie Liu and Henglin Liu and Gongye Liu and Jun Zheng and Wanyuan Pang and Ao Ma and Zhenyu Xie and Xintao Wang and Meng Wang and Pengfei Wan and Xiaodan Liang},
year={2025},
eprint={2510.22319},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2510.22319},
}
If you find Flow-DPO useful for your research or projects, we would greatly appreciate it if you could cite the following paper:
@article{liu2025improving,
title={Improving video generation with human feedback},
author={Liu, Jie and Liu, Gongye and Liang, Jiajun and Yuan, Ziyang and Liu, Xiaokun and Zheng, Mingwu and Wu, Xiele and Wang, Qiulin and Qin, Wenyu and Xia, Menghan and others},
journal={arXiv preprint arXiv:2501.13918},
year={2025}
}
