VINS-120K

# VINS-120K: Ultra High-Resolution Image Editing with A Large-Scale Dataset <p> <a href="https://nju-pcalab.github.io/projects/VINS-120K/"> <img src="https://img.shields.io/badge/Homepage-VINS--120K-2ea44f?style=for-the-badge" alt="Homepage"> </a> <a href="https://arxiv.org/abs/2605.23518"> <img src="https://img.shields.io/badge/arXiv-2605.23518-b31b1b?style=for-the-badge&logo=arxiv&logoColor=white" alt="arXiv"> </a> <a href="https://modelscope.cn/datasets/vivo/VINS-120K"> <img src="https://img.shields.io/badge/ModelScope-vivo%2FVINS--120K-624aff?style=for-the-badge" alt="ModelScope"> </a> <a href="https://huggingface.co/datasets/openvivo/VINS-120K"> <img src="https://img.shields.io/badge/Hugging%20Face-vivo%2FVINS--120K-yellow?style=for-the-badge&logo=huggingface" alt="Hugging Face"> </a> </p> ## Dataset Summary **VINS-120K** is a large-scale dataset for instruction-based ultra-high-resolution image editing. It contains **120K carefully curated editing triplets**, where each sample consists of a source image, an edited target image, and an English editing instruction. Each image is ultra-high-resolution, with resolution no lower than **4K**. The dataset covers **13 editing types** across local editing, global editing, camera movement, and personalized generation. VINS-120K is built from real-world ultra-high-resolution videos and high-quality open-source editing data. The data are filtered through multiple stages to ensure visual quality, instruction alignment, and aesthetic fidelity. <div align="center"> <img src="./assets/vins120k_overview.jpg" width="95%"> </div> <div align="center"> **Figure:** Overview and visualized examples of edited triplets in VINS-120K. </div> ## Dataset Structure Each annotation entry follows the format below: ```text edit_type: Editing category input: Source image relative path output: Edited image relative path instruction: English editing instruction ``` The training data are divided into three parts: ```text nano-consistent/ ultravideo/ x2edit/ ``` The corresponding annotation files are: ```text nano-consistent.json ultravideo.json x2edit.json ``` ## Data Files The repository contains both image archives and JSON annotation files. ```text . ├── nano-consistent/ │ ├── Image.tar.split.000 │ ├── Image.tar.split.001 │ ├── ... │ └── Image.tar.split.012 ├── ultravideo/ │ ├── clips_short_*.tar │ └── ... ├── x2edit/ │ ├── 0.tar.split.* │ ├── 1.tar.split.* │ ├── ... │ └── 7.tar.split.* ├── nano-consistent.json ├── ultravideo.json ├── x2edit.json ├── benchmark.tar └── assets/ └── vins120k_overview.jpg ``` ## Extracting the Training Data ### Extract `nano-consistent` `nano-consistent` contains one tar archive split into multiple parts: ```text Image.tar.split.000 Image.tar.split.001 ... Image.tar.split.012 ``` Run: ```bash mkdir -p extracted/nano-consistent cat nano-consistent/Image.tar.split.* > nano-consistent/Image.tar tar -xf nano-consistent/Image.tar -C extracted/nano-consistent ``` ### Extract `ultravideo` `ultravideo` contains independent tar archives and does not need merging. Run: ```bash mkdir -p extracted/ultravideo for tar_file in ultravideo/clips_short_*.tar; do echo "Extracting ${tar_file} ..." tar -xf "${tar_file}" -C extracted/ultravideo done ``` ### Extract `x2edit` `x2edit` contains multiple split tar archives. Each group should be merged separately: ```text 0.tar.split.* -> 0.tar 1.tar.split.* -> 1.tar ... 7.tar.split.* -> 7.tar ``` Run: ```bash mkdir -p extracted/x2edit for i in 0 1 2 3 4 5 6 7; do echo "Merging x2edit/${i}.tar.split.* ..." cat x2edit/${i}.tar.split.* > x2edit/${i}.tar echo "Extracting x2edit/${i}.tar ..." tar -xf x2edit/${i}.tar -C extracted/x2edit done ``` ## VINS-4KEval Benchmark We also provide `benchmark.tar`, which contains the **VINS-4KEval** benchmark introduced in our paper. VINS-4KEval is designed for evaluating instruction-based ultra-high-resolution image editing models. It contains **509 carefully selected 4K/UHR test samples** and covers the same 13 editing types as VINS-120K, including local editing, global editing, camera movement, and personalized generation. This benchmark is intended to evaluate model performance in ultra-high-resolution image editing, especially in terms of instruction following, structural preservation, and fine-detail quality. ### Extract VINS-4KEval ```bash mkdir -p extracted/benchmark tar -xf benchmark.tar -C extracted/benchmark ``` ## License This dataset is released under the **Apache License 2.0**. ## Citation If you find this dataset useful, please cite: ```bibtex @inproceedings{chen2026vins, title={VINS-120K: Ultra High-Resolution Image Editing with A Large-Scale Dataset}, author={Chen, Zhizhou and Guan, Shanyan and Gao, Zhanxin and Ci, En and Ge, Yanhao and Li, Wei and Zhang, Zhenyu and Yang, Jian and Tai, Ying}, booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, pages={15302--15312}, year={2026} } ```