--- pretty_name: SmoothStyle license: other task_categories: - image-to-image tags: - image - style-transfer - controllable-generation - strotss size_categories: - 10K **Release status:** the repository structure and data indices are prepared for reproducibility, but the dataset must not be made public until the remaining Style30K per-image rights review and project-generated style-image generation records have been completed. See [DATASET_LICENSE.md](DATASET_LICENSE.md) and [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md). ## Dataset structure The dataset contains 1,812 content-style pairs. Each pair has ten stylized targets corresponding to strengths 0.1 through 1.0. | Split | Pairs | Examples | | --- | ---: | ---: | | train | 1,449 | 14,490 | | test | 363 | 3,630 | | total | 1,812 | 18,120 | Each JSONL row is one training example: ```json { "id": "0000_1011_06", "pair_id": "0000_1011", "content_id": "0000", "style_id": "1011", "strength_id": 6, "strength": 0.6, "content_file_name": "content/0000.jpg", "style_file_name": "style/1011.jpg", "target_file_name": "target/s06/0000_1011_6.jpg", "style_source": "project-generated", "generator": "STROTSS" } ``` The `*_file_name` fields follow the Hugging Face ImageFolder metadata convention and are relative to the split directory containing `metadata.jsonl`. ```text data/ ├── train/ │ ├── metadata.jsonl │ ├── content/ │ ├── style/ │ └── target/s01 ... target/s10/ └── test/ ├── metadata.jsonl ├── content/ ├── style/ └── target/s01 ... target/s10/ ``` Strength IDs `1` through `10` map to scalar strengths `0.1` through `1.0`. The source collection also used IDs `0` and `11` for duplicated content and style references; those redundant copies are not included in this repository. ## Loading After publishing this repository on the Hugging Face Hub: ```python from datasets import load_dataset dataset = load_dataset("ORG/SmoothStyle") sample = dataset["train"][0] content = sample["content"] style = sample["style"] target = sample["target"] strength = sample["strength"] ``` For a local checkout: ```python dataset = load_dataset("imagefolder", data_dir="data") ``` ## Split policy The existing project split is preserved using seed 42 and an 80/20 split over complete `(content_id, style_id)` pairs. Pair IDs and content IDs do not overlap between train and test. Some style IDs occur in both splits; therefore this split evaluates unseen content-style pairs, not strictly unseen-style generalization. Exact pair lists are stored in `metadata/split_manifest.json`. ## Image sources Content images were randomly selected from the `content/` collection of [OmniStyle-150K](https://huggingface.co/datasets/StyleXX/OmniStyle-150k). Its Hugging Face Dataset Card declares the dataset under Apache-2.0. The OmniStyle paper describes its content images as generated using FLUX. SmoothStyle preserves this upstream declaration and grants no additional rights in the OmniStyle source material. Style IDs beginning with `0` were selected from Style30K. Style IDs beginning with `1` were generated by the SmoothStyle project. The complete per-asset status is recorded in `metadata/sources.jsonl`. According to the StyleTokenizer documentation, some Style30K images were sourced via LAION-5B. LAION-5B licenses its metadata under CC BY 4.0, while the underlying images remain subject to the rights of their original rightsholders. SmoothStyle does not infer that a specific Style30K image came from LAION-5B unless a reliable source mapping is available. ## Generation method Stylized target images were generated by the dataset authors using an implementation of Style Transfer by Relaxed Optimal Transport and Self-Similarity (STROTSS). Users should cite the STROTSS paper in addition to this dataset. The target generation operation does not remove rights associated with source content or style images. The rights status of every target therefore depends on the rights status of both source images. ## Intended uses SmoothStyle is intended for research on controllable image style transfer, continuous style-strength conditioning, image-to-image generation, and related evaluation methods. The dataset should not be used to claim unseen-style generalization with the provided split, to identify people, or to reproduce or impersonate living artists without appropriate permission. ## Licensing This is a mixed-rights dataset and is marked `license: other`. Project-authored metadata is offered under CC BY 4.0. No blanket license is granted for third-party images. See: - [DATASET_LICENSE.md](DATASET_LICENSE.md) - [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md) - [LICENSES/CC-BY-4.0.txt](LICENSES/CC-BY-4.0.txt) - [LICENSES/APACHE-2.0.txt](LICENSES/APACHE-2.0.txt) ## Reproducibility and integrity - `metadata/dataset_stats.json`: dataset counts and source categories - `metadata/split_manifest.json`: deterministic pair membership - `metadata/sources.jsonl`: per-asset provenance and rights status - `metadata/checksums.sha256`: SHA-256 checksums for generated repository data and metadata - `scripts/build_dataset.py`: deterministic build script - `scripts/validate_dataset.py`: structural, image-decoding, split, and checksum validation ## Citation Dataset citation information will be added when the accompanying SmoothStyle paper metadata is finalized. Please also cite OmniStyle, the source of the content-image subset: ```bibtex @inproceedings{wang2025omnistyle, title={OmniStyle: Filtering High Quality Style Transfer Data at Scale}, author={Wang, Ye and Liu, Ruiqi and Lin, Jiang and Liu, Fei and Yi, Zili and Wang, Yilin and Ma, Rui}, booktitle={CVPR}, year={2025} } ``` Please also cite STROTSS: ```bibtex @inproceedings{kolkin2019style, title={Style Transfer by Relaxed Optimal Transport and Self-Similarity}, author={Kolkin, Nicholas and Salavon, Jason and Shakhnarovich, Greg}, booktitle={CVPR}, year={2019} } ``` ## Contact and takedown Maintainer contact and a formal correction/takedown address must be added before public release. Until then, this repository should remain private.