This is the official implementation for the paper: HPSv3: Towards Wide-Spectrum Human Preference Score. First, we introduce a VLM-based preference model HPSv3, trained on a "wide spectrum" preference dataset HPDv3 with 1.08M text-image pairs and 1.17M annotated pairwise comparisons, covering both state-of-the-art and earlier generative models, as well as high- and low-quality real-world images. Second, we propose a novel reasoning approach for iterative image refinement, CoHP(Chain-of-Human-Preference), which efficiently improves image quality without requiring additional training data.

• [2025-08-19] 🖼️ We release DanceGRPO results of HPSv3! Great thanks to XueZeyue for training it!
• [2025-08-08] 🎉 We release HPDv3 dataset!.
• [2025-08-06] 🎉 We release HPSv3: inference code, training code, cohp code and HPSv3 model weights. And PyPI Package.

1. 🚀 Quick Start
2. 🌐 Gradio Demo
3. 🏋️ Training
4. 📊 Benchmark
5. 🎯 CoHP (Chain-of-Human-Preference)

HPSv3 is a state-of-the-art human preference score model for evaluating image quality and prompt alignment. It builds upon the Qwen2-VL architecture to provide accurate assessments of generated images.

# Method 1: Pypi download and install for inference.
pip install hpsv3

# Method 2: Install locally for development or training.
git clone https://github.com/MizzenAI/HPSv3.git
cd HPSv3

conda env create -f environment.yaml
conda activate hpsv3
# Recommend: Install flash-attn
pip install flash-attn==2.7.4.post1

pip install -e .

from hpsv3 import HPSv3RewardInferencer

# Initialize the model
inferencer = HPSv3RewardInferencer(device='cuda')

# Evaluate images
image_paths = ["assets/example1.png", "assets/example2.png"]
prompts = [
  "cute chibi anime cartoon fox, smiling wagging tail with a small cartoon heart above sticker",
  "cute chibi anime cartoon fox, smiling wagging tail with a small cartoon heart above sticker"
]

# Get preference scores
rewards = inferencer.reward(image_paths, prompts)
scores = [reward[0].item() for reward in rewards]  # Extract mu values
print(f"Image scores: {scores}")

Launch an interactive web interface to test HPSv3:

python gradio_demo/demo.py

The demo will be available at http://localhost:7860 and provides:

Human Preference Dataset v3 (HPD v3) comprises 1.08M text-image pairs and 1.17M annotated pairwise data. To modeling the wide spectrum of human preference, we introduce newest state-of-the-art generative models and high quality real photographs while maintaining old models and lower quality real images.

huggingface-cli download --repo-type dataset MizzenAI/HPDv3 --local-dir /your-local-dataset-path

Important Note: For simplicity, path1's image is always the prefered one

Important Notes: In HPDv3, we simply put the preferred sample at the first place (path1)

all.json contains all annotated pairs except for test.

[
    # samples from HPDv3 annotation pipeline 
    {
    "prompt": "Description of the visual content or the generation prompt.",
    "choice_dist": [12, 7],           # Distribution of votes from annotators (12 votes for image1, 7 votes for image2)
    "confidence": 0.9999907,         # Confidence score reflecting preference reliability, based on annotators' capabilities (independent of choice_dist)
    "path1": "images/uuid1.jpg",     # File path to the preferred image
    "path2": "images/uuid2.jpg",     # File path to the non-preferred image
    "model1": "flux",                # Model used to generate the preferred image (path1)
    "model2": "infinity"             # Model used to generate the non-preferred image (path2)
    },
    # samples from Midjourney
    {
    "prompt": "Description of the visual content or the generation prompt.",
    "choice_dist": null,             # No distribution of votes Information from Discord
    "confidence": null,              # No Confidence Information from Discord
    "path1": "images/uuid1.jpg",     # File path to the preferred image.
    "path2": "images/uuid2.jpg",     # File path to the non-preferred image.
    "model1": "midjourney",          # Comparsion between images generated from midjourney 
    "model2": "midjourney"           # Comparsion between images generated from midjourney 
    },
    # samples from Curated HPDv2
    {
    "prompt": "Description of the visual content or the generation prompt.",
    "choice_dist": null,              # No distribution of votes Information from the original HPDv2 traindataset
    "confidence": null,               # No Confidence Information from the original HPDv2 traindataset
    "path1": "images/uuid1.jpg",     # File path to the preferred image.
    "path2": "images/uuid2.jpg",     # File path to the non-preferred image.
    "model1": "hpdv2",          # No specific model name in the original HPDv2 traindataset, set to hpdv2 
    "model2": "hpdv2"           # No specific model name in the original HPDv2 traindataset, set to hpdv2 
    },
]

We sample part of training data from all.json to build training dataset train.json. Moreover, to improve robustness, we integrate random sampled part of data from Pick-a-pic and ImageRewardDB, which is pickapic.json and imagereward.json. For these two datasets, we only provide the pair infomation, and its corresponding image can be found in their official dataset repository.

[
    {
        "prompt": "Description of the visual content",
        "path1": "images/uuid1.jpg",     # Preferred sample
        "path2": "images/uuid2.jpg",     # Unpreferred sample
        "model1": "flux",                # Model used to generate the preferred sample (path1).
        "model2": "infinity",            # Model used to generate the non-preferred sample (path2).

    }
]

# Use Method 2 to install locally
git clone https://github.com/MizzenAI/HPSv3.git
cd HPSv3

conda env create -f environment.yaml
conda activate hpsv3
# Recommend: Install flash-attn
pip install flash-attn==2.7.4.post1

pip install -e .

# Train with 7B model
deepspeed hpsv3/train.py --config hpsv3/config/HPSv3_7B.yaml

To evaluate HPSv3 preference accuracy or human preference score of image generation model, follow the detail instruction is in Evaluate Insctruction

COHP is our novel reasoning approach for iterative image refinement that efficiently improves image quality without requiring additional training data. It works by generating images with multiple diffusion models, selecting the best one using reward models, and then iteratively refining it through image-to-image generation.

python hpsv3/cohp/run_cohp.py \
    --prompt "A beautiful sunset over mountains" \
    --index "sample_001" \
    --device "cuda:0" \
    --reward_model "hpsv3"

• --prompt: Text prompt for image generation (required)
• --index: Unique identifier for saving results (required)
• --device: GPU device to use (default: 'cuda:1')
• --reward_model: Reward model for scoring images
  • hpsv3: HPSv3 model (default, recommended)
  • hpsv2: HPSv2 model
  • imagereward: ImageReward model
  • pickscore: PickScore model

COHP uses multiple state-of-the-art diffusion models for initial generation: FLUX.1 dev, Kolors, Stable Diffusion 3 Medium, Playground v2.5

1. Multi-Model Generation: Generates images using all supported models
2. Reward Scoring: Evaluates each image using the specified reward model
3. Best Model Selection: Chooses the model that achieves the highest average score for its generated images
4. Iterative Refinement: Performs 4 rounds of image-to-image generation to improve quality
5. Adaptive Strength: Uses strength=0.8 for rounds 1-2, then 0.5 for rounds 3-4

We perform DanceGRPO as the reinforcement learning method. Here are some results. All experiments using the same setting and we use Stable Diffusion 1.4 as our backbone.

If you find HPSv3 useful in your research, please cite our work:

@misc{ma2025hpsv3widespectrumhumanpreference,
      title={HPSv3: Towards Wide-Spectrum Human Preference Score}, 
      author={Yuhang Ma and Xiaoshi Wu and Keqiang Sun and Hongsheng Li},
      year={2025},
      eprint={2508.03789},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.03789}, 
}

We would like to thank the VideoAlign codebase for providing valuable references.

For questions and support:

• Issues: GitHub Issues
• Email: yhshui@mizzen.ai & yhma@mizzen.ai