TST100K: A Large-Scale Triplet Dataset for Tone Style Transfer

Project Page | Paper

TST100K is a large-scale dataset for reference-based tone style transfer in photo retouching. The current release contains 108,808 image triplets constructed from the public PPR10K, MIT-Adobe FiveK, and Food-101 research datasets.

Each triplet contains a content image, a reference image, and a ground-truth target image. The task is to reproduce the photographic color and tone of the reference image while preserving the semantic content and spatial structure of the content image.

To improve supervision quality, TST100K applies both aesthetic quality filtering and tone-style consistency filtering. A learned tone style scorer removes reference-target pairs with inconsistent tone styles, while an aesthetic scorer filters stylizations that degrade the visual quality of the original image.

TST100K is released with our paper "Towards In-Context Tone Style Transfer with a Large-Scale Triplet Dataset."

For each triplet:

• Content is the input photo whose content and structure should be preserved.
• Reference provides the desired photographic color and tone.
• Ground truth is the content image retouched to match the reference tone style.

File Structure

Images are stored once and shared across triplets. To comply with Hugging Face repository limits, image files are sorted by filename and distributed sequentially with at most 10,000 files per shard. The relationships between files are recorded in triplet.json.

TST100K/
|-- content_images/
|   |-- 00/
|   |-- ...
|   `-- 01/
|-- style_images/
|   |-- 000/
|   |-- ...
|   `-- 016/
|-- triplet.json
`-- README.md

• content_images/ contains the input photographs in .jpg, .png, or .tif format.
• style_images/ contains preset-rendered reference and ground-truth images in .png format.
• triplet.json is an ordered JSON list. Each record stores paths relative to the TST100K root, including the shard directory.

An example manifest entry is:

{
  "content": "content_images/01/12615.tif",
  "reference": "style_images/011/10763_0527.png",
  "gt": "style_images/013/12615_0526.png"
}

The same image file may be referenced by multiple triplets. Use triplet.json as the authoritative pairing manifest; do not infer pairings or shard assignments from filenames.

Loading the Dataset

import json
from pathlib import Path

from PIL import Image

root = Path("/path/to/TST100K")
triplets = json.loads((root / "triplet.json").read_text(encoding="utf-8"))

sample = triplets[0]
content = Image.open(root / sample["content"])
reference = Image.open(root / sample["reference"])
ground_truth = Image.open(root / sample["gt"])

Data Sources and License

TST100K is constructed using images from the following public research datasets:

• PPR10K: a large-scale portrait photo retouching dataset. Please refer to the original repository for its license and terms of use.
• MIT-Adobe FiveK: also known as MIT5K. Its images are covered by LicenseAdobe.txt or LicenseAdobeMIT.txt.
• Food-101: a public dataset of food images. Please refer to the original dataset for its license and terms of use.

TST100K is released for non-commercial academic and research use only.

Commercial training, fine-tuning, product development, evaluation of commercial systems, resale, and use in commercial products or services are not allowed.

Users are responsible for complying with the licenses and terms of PPR10K, MIT-Adobe FiveK, Food-101, and any other applicable third-party materials. The TST100K usage terms do not replace or override the terms of the source datasets; where terms differ, the most restrictive applicable terms govern.

Citation

If you use TST100K in your research, please cite:

@misc{deng2026incontexttonestyletransfer,
  title         = {Towards In-Context Tone Style Transfer with A Large-Scale Triplet Dataset},
  author        = {Yuhai Deng and Huimin She and Wei Shen and Meng Li and Ruoxi Wu and Lunxi Yuan and Xiang Li},
  year          = {2026},
  eprint        = {2604.16114},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}