🎯 HPSv3: Towards Wide-Spectrum Human Preference Score (ICCV 2025)

Yuhang Ma^1,3*  Yunhao Shui^1,4*  Xiaoshi Wu²  Keqiang Sun^1,2†  Hongsheng Li^2,5,6†

¹Mizzen AI   ²CUHK MMLab   ³King’s College London   ⁴Shanghai Jiaotong University  

⁵Shanghai AI Laboratory   ⁶CPII, InnoHK  

^*Equal Contribution  ^†Equal Advising

📖 Introduction

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.

✨ Updates

• [2025-8-05] 🎉 We release HPSv3: inference code, training code, cohp code and model weights.

📑 Table of Contents

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

---

🚀 Quick Start

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.

💻 Installation

# 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 .

🛠️ Basic Usage

Simple Inference Example

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}")

---

🌐 Gradio Demo

Launch an interactive web interface to test HPSv3:

python gradio_demo/demo.py

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

🏋️ Training

📁 Dataset

Human Preference Dataset v3

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.

Detail information of HPD v3

Image Source	Type	Num Image	Prompt Source	Split
High Quality Image (HQI)	Real Image	57759	VLM Caption	Train & Test
MidJourney	-	331955	User	Train
CogView4	DiT	400	HQI+HPDv2+JourneyDB	Test
FLUX.1 dev	DiT	48927	HQI+HPDv2+JourneyDB	Train & Test
Infinity	Autoregressive	27061	HQI+HPDv2+JourneyDB	Train & Test
Kolors	DiT	49705	HQI+HPDv2+JourneyDB	Train & Test
HunyuanDiT	DiT	46133	HQI+HPDv2+JourneyDB	Train & Test
Stable Diffusion 3 Medium	DiT	49266	HQI+HPDv2+JourneyDB	Train & Test
Stable Diffusion XL	Diffusion	49025	HQI+HPDv2+JourneyDB	Train & Test
Pixart Sigma	Diffusion	400	HQI+HPDv2+JourneyDB	Test
Stable Diffusion 2	Diffusion	19124	HQI+JourneyDB	Train & Test
CogView2	Autoregressive	3823	HQI+JourneyDB	Train & Test
FuseDream	Diffusion	468	HQI+JourneyDB	Train & Test
VQ-Diffusion	Diffusion	18837	HQI+JourneyDB	Train & Test
Glide	Diffusion	19989	HQI+JourneyDB	Train & Test
Stable Diffusion 1.4	Diffusion	18596	HQI+JourneyDB	Train & Test
Stable Diffusion 1.1	Diffusion	19043	HQI+JourneyDB	Train & Test
Curated HPDv2	-	327763	-	Train

Download HPDv3

HPDv3 is comming soon! Stay tuned!

Pairwise Training Data Format

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

[
  {
    "prompt": "A beautiful landscape painting",
    "path1": "path/to/better/image.jpg",
    "path2": "path/to/worse/image.jpg",
    "confidence": 0.95
  },
  ...
]

🚀 Training Command

# 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

Important Config Argument

Configuration Section	Parameter	Value	Description
Model Configuration	rm_head_type	"ranknet"	Type of reward model head architecture
	lora_enable	False	Enable LoRA (Low-Rank Adaptation) for efficient fine-tuning. If False, language tower is fully trainable
	vision_lora	False	Apply LoRA specifically to vision components. If False, vision tower is fully trainable
	model_name_or_path	"Qwen/Qwen2-VL-7B-Instruct"	Path to the base model checkpoint
Data Configuration	confidence_threshold	0.95	Minimum confidence score for training data
	train_json_list	[example_train.json]	List of training data files
	test_json_list	[validation_sets]	List of validation datasets with names
	output_dim	2	Output dimension of the reward head for $\mu$ and $\sigma$
	loss_type	"uncertainty"	Loss function type for training

---

📊 Benchmark

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

Preference Accuracy of HPSv3

Model	ImageReward	Pickscore	HPDv2	HPDv3
CLIP ViT-H/14	57.1	60.8	65.1	48.6
Aesthetic Score Predictor	57.4	56.8	76.8	59.9
ImageReward	65.1	61.1	74.0	58.6
PickScore	61.6	70.5	79.8	65.6
HPS	61.2	66.7	77.6	63.8
HPSv2	65.7	63.8	83.3	65.3
MPS	67.5	63.1	83.5	64.3
HPSv3	66.8	72.8	85.4	76.9

Image Generation Benchmark of HPSv3

Model	Overall	Characters	Arts	Design	Architecture	Animals	Natural Scenery	Transportation	Products	Others	Plants	Food	Science
Kolors	10.55	11.79	10.47	9.87	10.82	10.60	9.89	10.68	10.93	10.50	10.63	11.06	9.51
Flux-dev	10.43	11.70	10.32	9.39	10.93	10.38	10.01	10.84	11.24	10.21	10.38	11.24	9.16
Playgroundv2.5	10.27	11.07	9.84	9.64	10.45	10.38	9.94	10.51	10.62	10.15	10.62	10.84	9.39
Infinity	10.26	11.17	9.95	9.43	10.36	9.27	10.11	10.36	10.59	10.08	10.30	10.59	9.62
CogView4	9.61	10.72	9.86	9.33	9.88	9.16	9.45	9.69	9.86	9.45	9.49	10.16	8.97
PixArt-Σ	9.37	10.08	9.07	8.41	9.83	8.86	8.87	9.44	9.57	9.52	9.73	10.35	8.58
Gemini 2.0 Flash	9.21	9.98	8.44	7.64	10.11	9.42	9.01	9.74	9.64	9.55	10.16	7.61	9.23
SDXL	8.20	8.67	7.63	7.53	8.57	8.18	7.76	8.65	8.85	8.32	8.43	8.78	7.29
HunyuanDiT	8.19	7.96	8.11	8.28	8.71	7.24	7.86	8.33	8.55	8.28	8.31	8.48	8.20
Stable Diffusion 3 Medium	5.31	6.70	5.98	5.15	5.25	4.09	5.24	4.25	5.71	5.84	6.01	5.71	4.58
SD2	-0.24	-0.34	-0.56	-1.35	-0.24	-0.54	-0.32	1.00	1.11	-0.01	-0.38	-0.38	-0.84

---

🎯 CoHP (Chain-of-Human-Preference)

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.

🚀 Usage

Basic Command

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

Parameters

• --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

Supported Generation Models

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

How COHP Works

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 produced the highest-scoring image
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

---

🦾 Results as Reward Model

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.

More Results of HPsv3 as Reward Model (Stable Diffusion 1.4)

📚 Citation

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

@inproceedings{hpsv3,
  title={HPSv3: Towards Wide-Spectrum Human Preference Score},
  author={Ma, Yuhang and Wu, Xiaoshi and Sun, Keqiang and Li, Hongsheng},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year={2025}
}

---

🙏 Acknowledgements

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

---

💬 Support

For questions and support:

• Issues: GitHub Issues
• Email: [email protected] & [email protected]