Dataset Card for StylExNet5k

StylExNet5k is a multi-style synthetic image dataset consisting of 5,000 images across 100 everyday object categories, each rendered in 10 distinct artistic or representational styles and placed in varied real-world contextual environments. It is intended to support evaluation and training of computer vision and vision–language models across style and context domains.

Dataset Details

Dataset Description

StylExNet5k is a synthetic dataset created using the Stable Diffusion XL (SDXL) base and refiner models. It contains 100 object categories, each rendered across 10 visual styles (e.g., photorealistic, watercolor, origami, pixel art, etc.) and various environmental contexts (e.g., indoors, outdoors). All images are square 1024×1024 PNGs, with additional resolutions (512, 384, 256, 128) provided for multi-scale use cases.

• Curated by: Bodhisatta Maiti
• Language(s) (NLP): English
• License: Creative Commons Attribution Non Commercial Share Alike 4.0 International

Dataset Sources

• Repository: https://huggingface.co/datasets/bodhisattamaiti/StylExNet5k
• Zenodo DOI: https://doi.org/10.5281/zenodo.15665050
• Kaggle DOI: https://doi.org/10.34740/kaggle/dsv/12167702

Uses

Direct Use

StylExNet5k is designed for the followinng, though this is not an exhaustive list:

1. Robustness testing of retrieval performance of vision–language models across stylistic variation and different resolution levels
2. Style-conditioned image retrieval and captioning
3. Zero-shot object detection and recognition across styles
4. Style-transfer benchmarking
5. Prompt faithfulness evaluation in diffusion models
6. Visual grounding and spatial reasoning research in stylized contexts
7. Style classification and clustering across different resolution levels
8. Few-shot classification of objects across styles

Usage Example

A sample example of the usage of the dataset can be seen in this notebook where a simple style classification has been done: https://www.kaggle.com/code/bodhisattamaiti/style-classification-on-stylexnet5k-data

Out-of-Scope Use

Commercial use is prohibited under the license.

The dataset should not be used for training real-world safety-critical systems without further validation.

Dataset Structure

StylExNet5k is a dataset consisting of 5000 high-resolution level AI-generated images of everyday objects placed in a variety of environments. Every object has been rendered in 10 different styles: photorealistic, oil painting, watercolor, sketch, voxel art, pixel art, origami, cyberpunk, isometric and papercut art. All the images have been generated using Stable Diffusion XL (Hugging Face: stabilityai/stable-diffusion-xl-base-1.0). The number of objects used is 100. Each object has 5 variants, and 10 styles have been applied, so, there are 50 images per object. The images generated by Stable Diffusion XL (SDXL) have the dimensions of 1024x1024, these can be found in the “images” folder of the dataset. From the original images of 1024x1024, other resolutions such as 512x512, 384x384, 256x256 and 128x128 have been derived and included in this dataset, and the resized images can be found in folders such as images_512, images_384, images_256 and images_128. All the resolution levels have 5000 images each. All the metadata related to the images such as object name, style, SDXL prompt, etc has been provided in a metadata csv file.

These are the column details related to the metadata csv file: object_id: There are 100 objects in the dataset, each object has been given unique id such as obj_001, obj_002,..obj_100. object_name: These are the names of the objects used in the dataset. variant: Each object and a particular style has 5 variants. for example, you will find 5 variants of Acoustic Guitar and Oil Painting style, variants are named as v1, v2,..v5. style: The style applied to a particular image, there are 10 styles as mentioned above. color: This is the color of the object. environment: This is the environment in which the object is placed or located. prompt: Prompt used to generate the image using SDXL. filename: filename of the image, named in this format:object_id_variant_style.png

Dataset Creation

Curation Rationale

The dataset was created to fill the gap in benchmarking datasets that assess robustness of vision and vision–language models across stylistic variation in real-world contexts. To the best of my knowledge, no such benchmark exists at this scale or style diversity.

Source Data

Data Collection and Processing

Images were generated using SDXL base and refiner pipelines with style-specific prompts and environment descriptions. High-quality prompts were manually designed and diversified.

Who are the source data producers?

All images were generated by a single curator using publicly available open-source tools (SDXL by Stability AI).

Annotations

Not applicable – all data is fully synthetic and self-labeled.

Personal and Sensitive Information

This dataset contains no personal, sensitive, or real-world identifying information.

Bias, Risks, and Limitations

Some stylistic representations may not align perfectly with prompts.

Artistic or cartoonish renderings may differ from natural scene statistics.

Real-world brand likeness of certain objects was manually filtered, but some ambiguity may persist.

Recommendations

The dataset is best suited for non-commercial research focused on style robustness, generative model evaluation, and zero-shot object recognition tasks.

Citation

BibTeX:

@misc{bodhisatta_maiti_2025, title={StylExNet5k}, url={https://www.kaggle.com/dsv/12167702}, DOI={10.34740/KAGGLE/DSV/12167702}, publisher={Kaggle}, author={Bodhisatta Maiti}, year={2025} }

APA:

Bodhisatta Maiti. (2025). StylExNet5k [Data set]. Kaggle. https://doi.org/10.34740/KAGGLE/DSV/12167702

Dataset Card Authors

Bodhisatta Maiti

Dataset Card Contact

[email protected]