WATER-Data: Datasets for WordArt-Oriented Scene Text Recognition

WATER-Data is the official dataset release for the paper "Advancing WordArt-Oriented Scene Text Recognition: Datasets and Methods" (ECCV 2026).

WordArt (artistic text) features highly customized fonts, textures, and layouts, making WAordArt-oriented scene TExt Recognition (WATER) substantially more challenging than general Scene Text Recognition (STR). The primary bottleneck for WATER is the lack of large-scale, stylistically diverse, and reliably annotated data. WATER-Data addresses this gap by providing a large-scale synthetic suite, a carefully deduplicated real training set, and a dedicated artistic-text benchmark.

• 📄 Paper (arXiv): https://arxiv.org/abs/2606.24484
• 💻 Code: https://github.com/YesianRohn/WATER
• 🧠 Model code (OpenOCR-WATERec): https://github.com/YesianRohn/OpenOCR-WATERec
• 🏋️ Model weights: https://huggingface.co/Yesianrohn/WATERec-Models
• 🖋️ Artistic fonts: https://huggingface.co/datasets/Yesianrohn/artistic-fonts
• 📝 WATER-Z captions: https://huggingface.co/datasets/Yesianrohn/WATER-Z_Captions

---

Dataset Overview

WATER-Data contains three components: a synthetic training suite (WATER-S), a real training set (WATER-R), and an artistic-text evaluation benchmark (WordArt-Bench).

Component	Subset	Role	#Instances	Source
WATER-S	WATER-T	Synthetic train	~1M	Tool-based rendering (SynthWordArt)
WATER-S	WATER-Z	Synthetic train	~1M	Generative model (Qwen3-VL + Z-Image)
WATER-R	–	Real train	3,225,130	Union14M-L + WordArt-Train + WAS-R (deduplicated)
WordArt-Bench	–	Evaluation	1,511	WordArt test split

All subsets are English WordArt in the current release.

---

Directory Structure

Every split is stored as a standalone LMDB database (data.mdb + lock.mdb), the format used by the OpenOCR framework.

WATER-Data/
├── WATER-R/                 # Real training set (~11.8 GB)
│   ├── data.mdb
│   └── lock.mdb
├── WATER-S/                 # Synthetic training suite
│   ├── WATER-T/             # Tool-rendered subset
│   │   ├── data.mdb
│   │   └── lock.mdb
│   └── WATER-Z/             # Model-generated subset
│       ├── data.mdb
│       └── lock.mdb
└── WordArt-Bench/           # Artistic-text benchmark (~325 MB)
    ├── data.mdb
    └── lock.mdb

---

Subset Details

WATER-S — Synthetic Suite (≈2M)

A 2M-scale synthetic artistic-text dataset, improving the scale of existing artistic text data by hundreds of times. It consists of two complementary subsets:

• WATER-T (Tool-based Rendering, ~1M). Generated with SynthWordArt, an artistic-text rendering engine built on SynthText / SynthTIGER. It replaces standard fonts with a library of 11,250 artistic fonts and adds rich layout patterns (curved lines, vertical text, multi-orientation layouts, perspective and stretching). It offers precise control over text content, font, and layout, with perfectly accurate labels.
• WATER-Z (Model-based Generation, ~1M). Generated by an automatic few-shot prompt-mining pipeline: Qwen3-VL-8B mines fine-grained captions (with an editable text placeholder) from real artistic text, expands them into 273,488 high-quality prompts, and Z-Image-Turbo synthesizes images at 256×256. It offers higher realism and diversity in background texture, layout composition, and global visual style.

WATER-T and WATER-Z are complementary: WATER-T provides strong controllability and label accuracy, while WATER-Z provides natural, design-like style diversity. Training on their combination covers both the "strongly controlled" and "style-diverse" regimes.

WATER-R — Real Training Set (3.2M)

A real-world training set re-constructed from three sources: Union14M-L, WordArt-Train, and WAS-R. Strict hashing deduplication is performed against all evaluation sets to avoid label leakage. It contains 3,225,130 text instances.

WordArt-Bench — Evaluation Benchmark

The artistic-text evaluation benchmark (test split of WordArt), with 1,511 images, used to report recognition accuracy. In the paper, our WATERec baseline reaches 90.40% accuracy on this benchmark — the first result to exceed 90% — surpassing both general-purpose and OCR-specialized vision-language models by a large margin.

---

Usage

Each LMDB database stores image–label pairs in the OpenOCR convention. A minimal reading example:

import lmdb

env = lmdb.open(
    "WATER-Data/WordArt-Bench",  # folder containing data.mdb / lock.mdb
    readonly=True, lock=False, readahead=False, meminit=False,
)

with env.begin(write=False) as txn:
    num_samples = int(txn.get(b"num-samples"))
    # keys follow the OpenOCR layout, e.g.:
    #   image-000000001 -> raw image bytes
    #   label-000000001 -> ground-truth text
    img_buf = txn.get(b"image-000000001")
    label = txn.get(b"label-000000001").decode("utf-8")

print(num_samples, label)

For training and evaluation, we recommend using the official framework OpenOCR-WATERec, which consumes these LMDB databases directly.

To download the dataset:

# Requires: pip install -U "huggingface_hub[cli]"
hf download Yesianrohn/WATER-Data --repo-type dataset --local-dir ./WATER-Data

---

Intended Use

WATER-Data is intended for research on scene text recognition, especially artistic / WordArt text. Typical uses include: training and benchmarking STR models, studying synthetic-data strategies (tool-based vs. generative), and evaluating general / OCR-specialized VLMs on challenging stylized text.

---

License

Released under the Apache 2.0 license. The dataset is built upon publicly available STR data sources (Union14M-L, WordArt, WAS-R) and synthetic content; please also respect the original licenses of these underlying datasets.

---

Citation

If you use WATER-Data in your research, please cite our paper:

@inproceedings{water2026eccv,
  title     = {Advancing WordArt-Oriented Scene Text Recognition: Datasets and Methods},
  author    = {Ye, Xingsong and Du, Yongkun and Zhang, Jiaxin and Zhang, Haojie and Sun, Chong and Li, Chen and Lyu, Jing and Chen, Zhineng},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}