StylExNet5k: A Multi-Style Visual Object Dataset with Environment Contexts

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.

The dataset has multiple use cases, though this is not an exhaustive list:

VLM image-text retrieval performance across styles and across different image resolution levels.

Style classification and clustering

Assessing the prompt faithfulness of diffusion models.

Assessing the style-conditioned image generation capabilities of diffusion models. 

Zero-shot object detection and segmentation capabilities of models across styles.