We introduce Adv-GRPO, an RL framework with an adversarial reward that iteratively updates both the reward model and the generator. Our method Adv-GRPO improves text-to-image (T2I) generation in three ways:

1. Alleviate Reward Hacking, achieving higher perceptual quality while maintaining comparable benchmark performance (e.g., PickScore, OCR), as shown in the top-left human evaluation panel;

2. Visual Foundation Model as Reward, leveraging visual foundation models (e.g., DINO) for rich visual priors, leading to overall improvements as shown in middle-top human evaluation results;

3. RL-based Distribution Transfer, enabling style customization by aligning generations with reference domains

2026-2-21

🎉 Our paper Adv-GRPO has been accepted to CVPR 2026 (Main Conference / Main Track)!

2025-12-15

• Released the checkpoints trained with OCR and GenEval prompts under the DINO reward framework:
  • DINO_OCR
  • DINO_GENEVAL

2025-11-27

• We also released our reference datasets.

  • QWen_Image_PickScore

• The Adv-GRPO demo is now available on Hugging Face:

  • PickScore
  • DINOv2

2025-11-25

• We release the code of Adv-GRPO training code, inference code and the pretrained ckpt.

• Release the reference dataset used in our work
• Release the DINO reward checkpoint trained with GenEval and OCR prompts
• Release the style transfer checkpoint
• Try more base models like QWen-Image

Clone this repository and install packages.

git clone https://github.com/showlab/Adv-GRPO.git
cd Adv-GRPO
conda create -n adv_grpo python=3.10.16 -y
pip install -e .

We use the Qwen-Image model (https://github.com/QwenLM/Qwen-Image) to generate reference images.

First, install the dependencies required by Qwen-Image.

python reference_imgs_scripts/qwen_generate_multi.py \
    --node_rank 0 \
    --num_nodes 1 \
    --num_variations 8 \
    --output_dir "" \
    --text_file ""

The reference images will be saved in output_dir and the json file will be like this:

{
  "middle-aged man with a beard giving a thumbs up, upper body, green fields in the background": [
    "node0_rank3_00000_0.png",
    "node0_rank3_00000_1.png",
    ...
  ],
  "king charles spaniel with planets for eyes, ethereal, midjourney style lighting and shadows, insanely detailed, 8k, photorealistic": [
    "node0_rank3_00001_0.png",
    "node0_rank3_00001_1.png",
    ...
  ],
  ...

}

And if you do not want to generate, you can use our generated images: 🤗QWen_PickScore

Some tips:

• Our reference dataset is relatively large — the full set is about 60 GB if you choose to download it.

• Actually, we do not use all images during training. Similarly, not all prompts are covered when using DINOv2.

• Based on our ablation studies, using a smaller subset of reference images and prompts can still achieve comparable performance to using the full dataset when observing the DINO similarity.

• If you prefer not to use our dataset or have a better alternative, you can use your own dataset and simply adapt it to the required format.

Firstly, we set the config file .config/grpo.py

def eval_sd3_fast():
    ...
     config.train.lora_path = ""
    config.save_folder = ""
    config.json_path = ""
    config.reference_image_path = ""
    config.test_reference_image_path = ""
    ...

• lora_path: LoRA checkpoint path
• save_folder: Output directory
• json_path: JSON metadata file where each key is a prompt and each value is a list containing the file paths of the corresponding reference images.
• reference_image_path: Reference images for inference (optional)
• test_reference_image_path: Test-time reference images (optional)

Secondly,

bash scripts/multi_node/sd3_fast/eval.sh 

You can modify the value after --prompts to try any text prompt you like.

If you want to generate one image,

python3 inference_t2i.py --config config/grpo.py:eval_sd3_fast --prompts "a flower on a planet"

You can modify the value after --prompts to try any text prompt you like.

The config file is in the .config/grpo.py

def dino_cotrain_sd3_patch_fast():
    ...
    config.json_path = ""
    config.refernce_image_path = ""
    config.test_reference_image_path = ""
    ...

We use deepspeed stage2 to save the memory.

# zero2
accelerate launch --config_file scripts/accelerate_configs/deepspeed_zero2.yaml
# zero3
accelerate launch --config_file scripts/accelerate_configs/deepspeed_zero3.yaml

Single-node training:

# sd3 grpo with DINO reward
bash scripts/grpo_dino.sh

# sd3 grpo with PickScore reward
bash scripts/grpo_pickscore.sh

1. You can adjust the parameters in config/grpo.py to tune different hyperparameters.

This repo is based on Flow-GRPO . We thank the authors for their valuable contributions to the AIGC community.

If you find Adv-GRPO useful for your research or projects, we would greatly appreciate it if you could cite the following paper:

@article{mao2025image,
  title={The Image as Its Own Reward: Reinforcement Learning with Adversarial Reward for Image Generation},
  author={Mao, Weijia and Chen, Hao and Yang, Zhenheng and Shou, Mike Zheng},
  journal={arXiv preprint arXiv:2511.20256},
  year={2025}
}