HPDv3

<div align="center">

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

[](https://research.nvidia.com/labs/par/addit/)
[](https://arxiv.org/abs/2508.03789)
[](https://arxiv.org/abs/2508.03789)
[](https://huggingface.co/MizzenAI/HPSv3)
[](https://github.com/MizzenAI/HPSv3)

**Yuhang Ma**^1,3*&ensp; **Yunhao Shui**^1,4*&ensp; **Xiaoshi Wu**²&ensp; **Keqiang Sun**^1,2†&ensp; **Hongsheng Li**^2,5,6†

¹Mizzen AI&ensp;&ensp; ²CUHK MMLab&ensp;&ensp; ³King’s College London&ensp;&ensp; ⁴Shanghai Jiaotong University&ensp;&ensp; ⁵Shanghai AI Laboratory&ensp;&ensp; ⁶CPII, InnoHK&ensp;&ensp;

^*Equal Contribution&ensp; ^†Equal Advising

</div>

<p align="center">
<img src="assets/teaser.png" alt="Teaser" width="900"/>
</p>

# 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.

## How to Use
```bash
cat images.tar.gz.* | gunzip | tar -xv
```

## Detail information of HPDv3

| 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 |

## Dataset Visualization
<p align="left">
<img src="assets/datasetvisual_0.jpg" alt="Dataset" width="900"/>
</p>

## Dataset Structure

### All Annotated Pairs (`all.json`)

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

`all.json` contains **all** annotated pairs except for test. There are three types of training samples in the json file.

```json
[
// 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
},
...
]
```

### Train set (`train.json`)
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](https://huggingface.co/datasets/pickapic-anonymous/pickapic_v1) and [ImageRewardDB](https://huggingface.co/datasets/zai-org/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.

### Test Set (`test.json`)
```json
[
{
"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).

},
...
]
```

<div align="center">

# 🎯 HPSv3: 面向宽谱人类偏好评分（ICCV 2025）

[](https://research.nvidia.com/labs/par/addit/)
[](https://arxiv.org/abs/2508.03789)
[](https://arxiv.org/abs/2508.03789)
[](https://huggingface.co/MizzenAI/HPSv3)
[](https://github.com/MizzenAI/HPSv3)

**马宇航**^1,3*&ensp; **水云浩**^1,4*&ensp; **吴小诗**²&ensp; **孙克强**^1,2†&ensp; **李洪生**^2,5,6†

¹Mizzen AI&ensp;&ensp; ²香港中文大学多媒体实验室（CUHK MMLab）&ensp;&ensp; ³伦敦国王学院&ensp;&ensp; ⁴上海交通大学&ensp;&ensp; ⁵上海人工智能实验室&ensp;&ensp; ⁶香港创新及科技伙伴计划中心（CPII, InnoHK）&ensp;&ensp;

^*共同第一作者&ensp; ^†共同通讯作者

</div>

<p align="center">
<img src="assets/teaser.png" alt="示例图" width="900"/>
</p>

# 人类偏好数据集v3

人类偏好数据集v3（Human Preference Dataset v3，简称HPD v3）包含108万组图文对与117万条带标注的成对数据。为建模宽谱人类偏好，我们在保留旧模型生成样本与低质量真实图像的基础上，引入了当前最先进的生成式模型与高质量真实摄影作品。

## 使用方法
bash
cat images.tar.gz.* | gunzip | tar -xv

## HPDv3详细信息

| 图像来源 | 类型 | 图像数量 | 提示词来源 | 数据集划分 |
|--------------|------|-----------|---------------|-------|
| 高质量图像（HQI） | 真实图像 | 57759 | 视觉语言模型标注标题 | 训练集与测试集 |
| MidJourney | - | 331955 | 用户 | 训练集 |
| CogView4 | DiT | 400 | HQI+HPDv2+JourneyDB | 测试集 |
| FLUX.1 dev | DiT | 48927 | HQI+HPDv2+JourneyDB | 训练集与测试集 |
| Infinity | 自回归模型 | 27061 | HQI+HPDv2+JourneyDB | 训练集与测试集 |
| Kolors | DiT | 49705 | HQI+HPDv2+JourneyDB | 训练集与测试集 |
| HunyuanDiT | DiT | 46133 | HQI+HPDv2+JourneyDB | 训练集与测试集 |
| Stable Diffusion 3 Medium | DiT | 49266 | HQI+HPDv2+JourneyDB | 训练集与测试集 |
| Stable Diffusion XL | 扩散模型 | 49025 | HQI+HPDv2+JourneyDB | 训练集与测试集 |
| Pixart Sigma | 扩散模型 | 400 | HQI+HPDv2+JourneyDB | 测试集 |
| Stable Diffusion 2 | 扩散模型 | 19124 | HQI+JourneyDB | 训练集与测试集 |
| CogView2 | 自回归模型 | 3823 | HQI+JourneyDB | 训练集与测试集 |
| FuseDream | 扩散模型 | 468 | HQI+JourneyDB | 训练集与测试集 |
| VQ-Diffusion | 扩散模型 | 18837 | HQI+JourneyDB | 训练集与测试集 |
| Glide | 扩散模型 | 19989 | HQI+JourneyDB | 训练集与测试集 |
| Stable Diffusion 1.4 | 扩散模型 | 18596 | HQI+JourneyDB | 训练集与测试集 |
| Stable Diffusion 1.1 | 扩散模型 | 19043 | HQI+JourneyDB | 训练集与测试集 |
| 精选HPDv2 | - | 327763 | - | 训练集 |

## 数据集可视化
<p align="left">
<img src="assets/datasetvisual_0.jpg" alt="数据集示例" width="900"/>
</p>

## 数据集结构

### 全量标注成对数据（all.json）

**重要说明：在HPDv3中，我们将偏好样本置于首位（即path1）**

`all.json`包含除测试集外的所有标注成对数据，该文件内包含三类训练样本。

json
[
// 来自HPDv3标注流水线的样本
{
"prompt": "视觉内容描述或生成提示词。",
"choice_dist": [12, 7], // 标注者的投票分布（12票支持图像1，7票支持图像2）
"confidence": 0.9999907, // 置信度分数，反映偏好的可靠性，基于标注者的能力水平（与choice_dist无关）
"path1": "images/uuid1.jpg", // 偏好图像的文件路径
"path2": "images/uuid2.jpg", // 非偏好图像的文件路径
"model1": "flux", // 生成偏好图像的模型（path1对应的模型）
"model2": "infinity" // 生成非偏好图像的模型（path2对应的模型）
},
// 来自Midjourney的样本
{
"prompt": "视觉内容描述或生成提示词。",
"choice_dist": null, // 无投票分布信息（来自Discord的原始数据）
"confidence": null, // 无置信度信息（来自Discord的原始数据）
"path1": "images/uuid1.jpg", // 偏好图像的文件路径。
"path2": "images/uuid2.jpg", // 非偏好图像的文件路径。
"model1": "midjourney", // 对比Midjourney生成的图像
"model2": "midjourney" // 对比Midjourney生成的图像
},
// 来自精选HPDv2的样本
{
"prompt": "视觉内容描述或生成提示词。",
"choice_dist": null, // 无投票分布信息（来自原始HPDv2训练集）
"confidence": null, // 无置信度信息（来自原始HPDv2训练集）
"path1": "images/uuid1.jpg", // 偏好图像的文件路径。
"path2": "images/uuid2.jpg", // 非偏好图像的文件路径。
"model1": "hpdv2", // 原始HPDv2训练集无特定模型名称，故设为hpdv2
"model2": "hpdv2" // 原始HPDv2训练集无特定模型名称，故设为hpdv2
},
...
]

### 训练集（train.json）
我们从`all.json`中采样部分训练数据以构建`train.json`训练集。此外，为提升模型鲁棒性，我们还集成了从[Pick-a-pic](https://huggingface.co/datasets/pickapic-anonymous/pickapic_v1)与[ImageRewardDB](https://huggingface.co/datasets/zai-org/ImageRewardDB)中随机采样的部分数据，对应的数据文件为`pickapic.json`与`imagereward.json`。对于这两个数据集，我们仅提供成对数据的信息，其对应的图像可在其官方数据集仓库中获取。

### 测试集（test.json）
json
[
{
"prompt": "视觉内容描述",
"path1": "images/uuid1.jpg", // 偏好样本
"path2": "images/uuid2.jpg", // 非偏好样本
"model1": "flux", // 生成偏好样本的模型（path1对应的模型）
"model2": "infinity", // 生成非偏好样本的模型（path2对应的模型）

},
...
]