File size: 5,877 Bytes
4e8eb2c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0d96c79
4e8eb2c
0d96c79
4e8eb2c
 
 
0d96c79
 
 
 
 
 
 
 
 
 
 
 
 
4e8eb2c
 
0d96c79
 
 
4e8eb2c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0d96c79
4e8eb2c
 
 
 
 
 
 
 
 
 
 
0d96c79
 
4e8eb2c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0d96c79
 
 
 
 
4e8eb2c
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
---
license: apache-2.0
task_categories:
- image-classification
- text-to-image
language:
- en
- zh
tags:
- preference
- reward-model
- text-to-image
- human-feedback
pretty_name: HPDv3++
size_categories:
- 100K<n<1M
---

# HPDv3++: A Dual-Dimension Preference Dataset for Text-to-Image Reward Modeling

HPDv3++ is a large-scale human-preference dataset for text-to-image (T2I) generation, built on a frontier generator (Qwen-Image) and annotated along two axes: **text-following (TF)** and **aesthetic quality (Aes)**. It is the dataset used to train [HPSv3++](https://huggingface.co/Junjun2333/HPSv3-PlusPlus), a capability-aware and RL-iteration-aware reward model.

Each preference pair stores a **preferred** image (`path1`) and a **non-preferred** image (`path2`) for the same prompt.

## Quick start

```bash
pip install -U "huggingface_hub[cli]"
hf download Junjun2333/HPDv3-PlusPlus --repo-type dataset --local-dir HPDv3pp
cd HPDv3pp
# Reassemble and extract our image pool (split tar parts -> images/qwen_image, images/rollout, images/thumbs):
cat images.tar.part* | tar -xf -
```

The split tar contains **only the images we generated** (`images/qwen_image/`,
`images/rollout/`, `images/thumbs/`). The `stage1_ref.json` reference pairs point
to the **original HPDv3 images** (`images/hpdv3/...`), which we do **not**
re-host here. If you need them (only required to reproduce HPSv3++ Stage 1 with
the original HPDv3 reference set), download the HPDv3 images from the official
repo and place them under `images/hpdv3/`:

```bash
# Original HPDv3 images (only needed for stage1_ref.json)
hf download MizzenAI/HPDv3 --repo-type dataset --include "images.tar.gz.*" --local-dir hpdv3_src
cat hpdv3_src/images.tar.gz.* | gunzip | tar -xv   # then move/symlink the resulting images into images/hpdv3/
```

After extraction you get an `images/` directory. Every path in the JSON files
(`path1` / `path2` / `image_path`) is **relative** and resolves against the
repo root, e.g. `images/qwen_image/prompt_000000/6.jpg`. The four ready-to-use
train/test files (`train_aes`, `train_tf`, `test_aes`, `test_tf`) reference only
our own images and need no HPDv3 download.

## What you can use directly

These four files are ready-to-use, self-contained, and **do not require any
HPSv3++ code or model** -- just images + JSON. Each record is
`{"path1": <preferred>, "path2": <non-preferred>, "prompt": <text>}` (the same
format as HPSv3/HPDv3), with `path1` preferred over `path2`.

| File | Pairs | Use |
|---|---|---|
| `train/train_aes.json` | 100,463 | Training -- aesthetic preference |
| `train/train_tf.json`  | 90,908  | Training -- text-following preference |
| `test/test_aes.json`   | 5,720   | Evaluation -- aesthetic |
| `test/test_tf.json`    | 4,465   | Evaluation -- text-following |

The training and test sets are **disjoint** (no shared pairs), including across
the two axes (aes/tf), so they can be used together without leakage.

## Repository layout

```
HPDv3-PlusPlus/
|-- images.tar.part00, images.tar.part01, ...   # split tar of OUR images (~268 GB; qwen_image + rollout + thumbs)
|-- train/
|   |-- train_aes.json        # 100,463  ready-to-use aesthetic training pairs
|   |-- train_tf.json         # 90,908   ready-to-use text-following training pairs
|   |-- stage1_labeled.json   # 191,466  labeled pairs (used by HPSv3++ Stage 1)
|   |-- stage1_ref.json       # 284,974  original HPDv3 reference pairs (Stage 1 OGD anti-forgetting)
|   |-- stage2_labeled.json   # 111,650  labeled pairs (used by HPSv3++ Stage 2)
|   |-- rollout.json          # 322,452  unlabeled rollouts, long format, one image per row
|   `-- ogd_std.json          # 58,242   pre-computed per-group std (also embedded in rollout.json)
|-- test/
|   |-- test_aes.json         # 5,720   ready-to-use aesthetic test pairs
|   `-- test_tf.json          # 4,465   ready-to-use text-following test pairs
`-- images/                   # after extraction: qwen_image/, rollout/, thumbs/ (ours);
                              #   hpdv3/ must be downloaded separately from MizzenAI/HPDv3 (only for stage1_ref)
```

## JSON formats

**Preference pairs** (`train_aes`, `train_tf`, `stage1_labeled`, `stage1_ref`, `stage2_labeled`, `test_aes`, `test_tf`):

| Field | Meaning |
|---|---|
| `path1` / `path2` | Preferred / non-preferred image (relative `images/...` path) |
| `prompt` | Text prompt |
| `choice_dist` / `confidence` / `model1` / `model2` | (where annotated) vote distribution, confidence, generator names; `null` otherwise. The ready-to-use `train_aes`/`train_tf` and `test` files keep only `path1/path2/prompt`. |

**`rollout.json`** (unlabeled rollouts for HPSv3++ Stage 2; long format, one image per row):

| Field | Meaning |
|---|---|
| `group_id` | Group id (same prompt + tier + iter_step form one group) |
| `source` | `capability` or `iteration` |
| `prompt` | Text prompt |
| `tier` | Generator tier |
| `iter_step` / `iter_norm` | Raw / normalized RL iteration |
| `capability` / `level` | Continuous capability score / discrete level |
| `image_path` | Relative image path |
| `ogd_std` | Pre-computed per-group std |

## Notes

- The images we host here (`qwen_image` + `rollout` + `thumbs`) are ~268 GB. The
  original HPDv3 images (`hpdv3/`, ~60 GB, referenced only by `stage1_ref.json`)
  are not re-hosted -- download them from MizzenAI/HPDv3 if needed (see Quick start).
- The ready-to-use train/test files reference only our own images, so they work
  with just the split tar above (no HPDv3 download needed).
- For the full two-stage training / evaluation pipeline (which additionally uses
  `rollout.json`, `stage1_ref.json`, etc.), see the HPSv3++ code repository.

## Citation

```bibtex
@misc{hpsv3pp,
  title  = {HPSv3++: Scaling Reward Models Across the Full Spectrum of Diffusion Model Capabilities},
  author = {HPSv3++ Team},
  year   = {2026}
}
```