csd100

## Introduction

Disentangling content and style from a single image, known as content-style decomposition (CSD), enables recontextualization of extracted content and stylization of extracted styles, offering greater creative flexibility in visual synthesis. While existing datasets focus on either style transfer or content preservation, they do not fully meet the requirements for evaluating CSD, prompting us to introduce CSD-100, a dataset of 100 images designed specifically for this task.

## Description

This dataset contains 100 unique samples covering wide range of content-style types such as animals, robots musical instruments, fruits, furniture, and so on.

## Sample Images


## Dataset Details

| | |
| --- | --- |
| Size | 7 MB |
| Train Size | N/A |
| Test Size | 100 images |
| Validation Size | N/A |
| Input Sample | Image |
| Label(s) | Synthesized images |

## Dataset Collection Process
* Starting with 400 content and 100 style prompts from RB-Modulation, we filter out ambiguous content terms, keeping 180 content and all style concepts.

* Using Flux-Schnell, we generate ~18,000 `<content> in <style>` images then manually select 1,000 representative images, further refined to 100 high-quality samples

## Data Format

* The archive contains 100 directories whose name have a format {content}+{style}.

* Each directory corresponds to one test sample and contains a single JPG image with a height of 1024 px and a width of 1024px.

* The filenames of the JPGs have the same name 00.jpg

## Dataset license

This dataset is intended for research purposes only.

Data License Agreement - Research Use – link to license agreement

## Dataset Citation Instructions

Please cite our paper if you use this dataset in your research.

```
@article{nguyen2025csd,
title = {CSD-VAR: Content-Style Decomposition in Visual Autoregressive Models},
author = {Nguyen, Quang-Binh and Luu, Minh and Nguyen, Quang and Tran, Anh and Nguyen, Khoi},
journal = {arXiv preprint arXiv:2507.13984},
year = {2025}
}
```

**Qualcomm AI Research**

At Qualcomm AI Research, we are advancing AI to make its core capabilities – perception, reasoning, and action – ubiquitous across devices. Our mission is to make breakthroughs in fundamental AI research and scale them across industries. By bringing together some of the best minds in the field, we’re pushing the boundaries of what’s possible and shaping the future of AI.

Qualcomm AI Research continues to invest in and support deep-learning research in computer vision. The publication of this dataset for use by the AI research community is one of our many initiatives.

Find out more about [Qualcomm AI Research](https://developer.qualcomm.com/forums/software/ai-research-datasets).

For any questions or technical support, please contact us at [e-mail]

*Qualcomm AI Research is an initiative of Qualcomm Technologies, Inc.*

## 引言
从单张图像中解耦内容与风格，即内容风格解耦（content-style decomposition, CSD），可支持对提取出的内容进行重上下文适配，以及对提取的风格进行风格化处理，为视觉合成提供更强的创作灵活性。现有数据集要么专注于风格迁移，要么聚焦内容保留，均无法完全满足内容风格解耦任务的评估需求，为此我们推出了专为该任务设计的CSD-100数据集，包含100张图像。

## 数据集概述
本数据集包含100个独特样本，涵盖动物、机器人、乐器、水果、家具等多种内容风格类型。

## 示例图像

## 数据集详情
| 指标 | 详情 |
| --- | --- |
| 数据集大小 | 7 MB |
| 训练集规模 | 无 |
| 测试集规模 | 100张图像 |
| 验证集规模 | 无 |
| 输入样本类型 | 图像 |
| 标签 | 合成图像 |

## 数据集采集流程
* 从RB-Modulation中获取400条内容提示词与100条风格提示词，过滤歧义性内容术语，最终保留180条内容概念与全部风格概念。
* 借助Flux-Schnell生成约18000张「<内容> in <风格>」格式的图像，随后人工筛选出1000张具有代表性的样本，进一步精调得到100张高质量样本。

## 数据格式
* 压缩包内包含100个目录，目录名称格式为`{content}+{style}`。
* 每个目录对应一个测试样本，内含一张尺寸为1024×1024像素的JPG图像。
* 所有JPG图像的文件名均为`00.jpg`。

## 数据集许可
本数据集仅用于研究目的。
数据使用许可协议——研究用途——[许可协议链接](https://developer.qualcomm.com/forums/software/ai-research-datasets)

## 数据集引用说明
若您在研究中使用本数据集，请引用我们的论文：

@article{nguyen2025csd,
title = {CSD-VAR: Content-Style Decomposition in Visual Autoregressive Models},
author = {Nguyen, Quang-Binh and Luu, Minh and Nguyen, Quang and Tran, Anh and Nguyen, Khoi},
journal = {arXiv preprint arXiv:2507.13984},
year = {2025}
}

**高通AI研究院（Qualcomm AI Research）**

高通AI研究院致力于推进人工智能技术发展，让感知、推理与行动等核心AI能力遍及各类设备。我们的使命是在基础AI研究领域取得突破，并将其规模化应用于各行业。通过汇聚该领域顶尖人才，我们不断突破技术边界，塑造人工智能的未来。

高通AI研究院持续投入并支持计算机视觉领域的深度学习研究。本次面向AI研究社区发布本数据集，便是我们众多举措之一。

了解更多[高通AI研究院](https://developer.qualcomm.com/forums/software/ai-research-datasets)相关信息。

如有任何疑问或技术支持需求，请发送邮件至`[e-mail]`与我们联系。

*高通AI研究院是高通技术公司（Qualcomm Technologies, Inc.）旗下的专项研究计划。*