README.md

---
configs:
- config_name: default
  data_files:
  - split: luminance
    path: "quantification_luminance.json"
  - split: chrominance
    path: "quantification_chrominance.json"
license: cc-by-nc-sa-4.0
---
# DocQT - Improving Document Forgery Localization Robustness via Diverse JPEG Quantization-Tables

Repository containing JPEG quantization tables used for the Real-QT protocol.
Dataset name: DocQT.
Only header-extracted quantization matrices are provided.

## Paper

- arXiv: https://arxiv.org/abs/2605.19688

## Hugging Face Dataset

- Dataset on Hugging Face Hub: https://huggingface.co/datasets/Kyliroco/DocQT

## Citation

If you use DocQT, this quantization-table repository, or build upon our article, please cite our paper:

```bibtex
@misc{ronfleuxcorail2026docqt,
	title={DocQT: Improving Document Forgery Localization Robustness via Diverse JPEG Quantization Tables},
	author={Kylian Ronfleux-Corail and Guillaume Bernard and Mickael Coustaty and Nicolas Sidere},
	year={2026},
	eprint={2605.19688},
	archivePrefix={arXiv},
	primaryClass={cs.CV},
	doi={10.48550/arXiv.2605.19688},
	url={https://arxiv.org/abs/2605.19688}
}
```

## Available quantization tables

- Luminance tables: `quantification_luminance.json`
  - Number of tables: 859
- Chrominance tables: `quantification_chrominance.json`
  - Number of tables: 294

## File format

Both files are stored in JSON for better portability and long-term compatibility.

- Root object: a list of quantization tables
- One table: a flat list of 64 integers
- Table values: integers

In other words, each file follows this structure:

- list[table]
- table = list[64 integer values]

## Example: load DocQT from Hugging Face Hub

```python
from datasets import load_dataset


def load_docqt_from_hub() -> tuple[list[list[int]], list[list[int]]]:
	dataset = load_dataset("Kyliroco/DocQT")

	luminance_split = dataset["luminance"]
	chrominance_split = dataset["chrominance"]

	# The 64-value quantization tables are stored in the "text" column.
	luminance_tables = luminance_split["text"]
	chrominance_tables = chrominance_split["text"]

	return luminance_tables, chrominance_tables


luminance_tables, chrominance_tables = load_docqt_from_hub()
```

Each selected table must be a flat list of 64 integer values before passing
it to Pillow through `qtables`.

## Example: use quantization tables with Pillow JPEG compression

```python
import json
from pathlib import Path
from PIL import Image


def load_quantization_tables(base_dir: str = ".") -> tuple[list, list]:
	base_path = Path(base_dir)

	luminance_tables = json.loads(
		(base_path / "quantification_luminance.json").read_text(encoding="utf-8")
	)
	chrominance_tables = json.loads(
		(base_path / "quantification_chrominance.json").read_text(encoding="utf-8")
	)

	return luminance_tables, chrominance_tables

luminance_tables, chrominance_tables = load_quantization_tables(".")

# Choose the table indices you want to use for compression.
luma_idx = 0
chroma_idx = 0

luma_qtable = luminance_tables[luma_idx]
chroma_qtable = chrominance_tables[chroma_idx]

with Image.open("input.png") as image:
	image = image.convert("RGB")
	image.save(
		"output_custom_qtables.jpg",
		format="JPEG",
		qtables=[luma_qtable, chroma_qtable],
		subsampling="4:2:0",
		optimize=True,
	)

# Note: avoid passing a quality value if you want to keep your custom qtables as-is.
```