RIROCommon

A scene text editing benchmark with equal-length and variable-length targets.

Benchmark 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

RIROCommon adapts six established scene text recognition benchmarks for scene text editing. Each source image is associated with an equal-length replacement and a variable-length target, enabling evaluation of content changes under both settings. It provides source crops, their original transcriptions, and target strings for editors to render.

The paper evaluates generated text with SVTRv2 using normalized word accuracy: non-vocabulary characters and spaces are removed, and text is lowercased before scoring. See the paper for the complete evaluation protocol.

Repository structure

The repository contains 12 LMDB databases: six source datasets in each editing setting.

Setting	Directory	Description
Equal-length editing	Equal/	Replacement text matches the source text length
Variable-length editing	Variable/	Replacement text is not constrained to match the source text length

Both directories contain these subsets:

Directory name	Benchmark name used in the paper
IIIT5k	IIIT
SVT	SVT
IC13	IC13
IC15	IC15
SVTP	SVTP
CUTE80	CUTE

Each subset directory, such as Equal/IC13/, contains data.mdb and lock.mdb.

Data format

The official benchmark reader uses the following LMDB keys. Sample indices start at 1 and use nine-digit zero padding.

Key	Value
num-samples	Number of samples in the database
image-000000001	Source image bytes
label-000000001	Original transcription, encoded as UTF-8
editlabel-000000001	Target text, encoded as UTF-8

Download

python -m pip install huggingface_hub

from huggingface_hub import snapshot_download

dataset_dir = snapshot_download(
    repo_id="Yesianrohn/RIROCommon",
    repo_type="dataset",
    local_dir="RIROCommon",
)

To download only one subset, add allow_patterns=["Equal/IC13/*"] to the call.

Run TextRIRO on the benchmark

First follow the TextRIRO installation and checkpoint instructions. Place the downloaded RIROCommon/ directory in the code repository root, or adjust --input_lmdb to its actual location. Run:

python inference_RIROCommon.py \
  --config configs/train.yaml \
  --ckpt_path textriro.pth \
  --input_lmdb RIROCommon/Equal/IC13 \
  --output_lmdb RIROOut/Equal/IC13 \
  --num_inference_steps 20 \
  --guidance_scale 2.2 \
  --batch_size 32 \
  --seed 42

Repeat for the desired subsets in Equal/ and Variable/, using a separate output directory for each. This command generates edited-image LMDBs; metric computation follows the paper's evaluation protocol.

For recognition training data generated with TextRIRO, see TextRIRO-3M.

Citation

If you use RIROCommon 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}
}