---
dataset_info:
  features:
  - name: image
    dtype: image
  - name: image_path
    dtype: string
  - name: caption
    dtype: string
  - name: merge_bbox
    list:
    - name: bbox
      sequence: int64
    - name: text
      dtype: string
  splits:
  - name: train
    num_bytes: 7981174979729.586
    num_examples: 4097983
  - name: test
    num_bytes: 449181829.0
    num_examples: 1000
  download_size: 1024574096356
  dataset_size: 7981624161558.586
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
---



<h2 align="center">TextGround4M: A Prompt-Aligned Dataset for Layout-Aware Text Rendering</h2>

<h5 align="center">
TextGround4M is a large-scale dataset for prompt-grounded, layout-aware text rendering in text-to-image (T2I) generation, introduced in our AAAI 2026 paper.
</h5>

<p align="center">
  <a href="https://arxiv.org/abs/2604.24459">
    <img src="https://img.shields.io/badge/arXiv-2604.24459-B31B1B?style=for-the-badge&amp;logo=arxiv&amp;logoColor=white" alt="arXiv" />
  </a>
  <a href="https://dongxingmao.github.io/TextGround4M.github.io/">
    <img src="https://img.shields.io/badge/Project_Page-TextGround4M-7289DA?style=for-the-badge&amp;logo=githubpages&amp;logoColor=white" alt="Project Page" />
  </a>
  <a href="https://github.com/CSU-JPG/TextGround4M">
    <img src="https://img.shields.io/badge/Code-TextGround4M-100000?style=for-the-badge&amp;logo=github&amp;logoColor=white" alt="Code" />
  </a>
</p>


## Dataset Summary

TextGround4M contains **4.1 million** prompt-image pairs, each annotated with:
- A natural language caption where all rendered text spans are explicitly quoted
- Span-level bounding boxes linking each quoted text to its spatial location in the image

This fine-grained annotation enables layout-aware, prompt-grounded supervision for T2I models — a capability missing from prior datasets like MARIO-10M and AnyWord-3M.

---

## Dataset Structure

### Splits

| Split | Samples | Description |
|---|---|---|
| `train` | ~4.1M | Training set with prompt-grounded bbox annotations |
| `test` | 1,000 | TextGroundEval benchmark (Easy / Medium / Hard) |

### Data Fields

| Field | Type | Description |
|---|---|---|
| `image` | `Image` | RGB image |
| `image_path` | `string` | Original filename (UUID hex) |
| `caption` | `string` | Natural language prompt with quoted text spans |
| `merge_bbox` | `list` | List of `{"bbox": [x1, y1, x2, y2], "text": "..."}` |

The `test` split additionally includes a `test/annotations.jsonl` file with `data_type` field (`easy` / `medium` / `hard`) for each sample.

---

## Usage

### Load with `datasets`

```python
from datasets import load_dataset

# Full dataset
ds = load_dataset("CSU-JPG/Textground4M")

# Train only
train = load_dataset("CSU-JPG/Textground4M", split="train")

# Test benchmark only
test = load_dataset("CSU-JPG/Textground4M", split="test")
```

### Load test split with `data_type` annotation

```python
from huggingface_hub import hf_hub_download
import json

path = hf_hub_download("CSU-JPG/Textground4M", "test/metadata.jsonl", repo_type="dataset")
records = [json.loads(l) for l in open(path)]
# Each record has: image_path, caption, merge_bbox, data_type
```

---


## License

This dataset is released under the [MIT License](https://opensource.org/licenses/MIT).  
Please also comply with the licenses of the original source datasets used in construction.

---

## Citation

```bibtex
@article{Mao_2026,
   title={TextGround4M: A Prompt-Aligned Dataset for Layout-Aware Text Rendering},
   volume={40},
   ISSN={2159-5399},
   url={http://dx.doi.org/10.1609/aaai.v40i10.37736},
   DOI={10.1609/aaai.v40i10.37736},
   number={10},
   journal={Proceedings of the AAAI Conference on Artificial Intelligence},
   publisher={Association for the Advancement of Artificial Intelligence (AAAI)},
   author={Mao, Dongxing and Wang, Yilin and Li, Linjie and Yang, Zhengyuan and Wang, Alex Jinpeng},
   year={2026},
   month=Mar, pages={7918–7926} }
```