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:

@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

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

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.