TextRIRO-3M

A 3-million-image dataset for long scene text recognition.

Dataset accompanying Real In, Real Out: What If We Only Use Real Data for Scene Text Editing?, accepted at NeurIPS 2026.

HF Paper · arXiv · NeurIPS 2026 · Code · Model · Demo · TextRIRO-3M · RIROCommon

Overview

TextRIRO-3M supplies synthetic training images for scene text recognition, with an emphasis on long and multi-word text. Starting from real U14M-Filter images, TextRIRO creates edits with different content and matching visual styles. An STR model filters edits for text correctness, and retained edits are concatenated with their original images to produce longer examples.

The default construction combines two equal-length strings, yielding even-length labels of at least eight characters. The paper separately studies asymmetric concatenation for broader length coverage.

Repository contents

The dataset is distributed as an LMDB database at the repository root.

File	Description
data.mdb	Dataset database
lock.mdb	LMDB lock file
README.md	Dataset card

Download

Install the Hugging Face Hub client:

python -m pip install huggingface_hub

Download the dataset files:

from huggingface_hub import snapshot_download

dataset_dir = snapshot_download(
    repo_id="Yesianrohn/TextRIRO-3M",
    repo_type="dataset",
    local_dir="TextRIRO-3M",
)

The database is stored directly in TextRIRO-3M/. Use an LMDB-compatible STR data pipeline to consume the downloaded data. The release provides database files rather than a Hugging Face datasets loading script.

Intended use

• Training or augmenting scene text recognition models.
• Studying recognition of long text and multiple words.
• Investigating the usefulness of scene text editing as a source of recognition training data.

TextRIRO-3M is the recognition training dataset associated with this work. For scene text editing evaluation with fixed target strings, use RIROCommon. For model setup and experimental protocols, see the code repository and paper.

Citation

If you use TextRIRO-3M in your research, please cite:

@inproceedings{ye2026textriro,
  title     = {Real In, Real Out: What If We Only Use Real Data for Scene Text Editing?},
  author    = {Ye, Xingsong and Du, Yongkun and Zhang, Jiaxin and Sun, Chong and Chen, Zhineng},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026},
  url       = {https://neurips.cc/virtual/2026/loc/sydney/poster/153296}
}