StyleCruxGen: A Visual Dataset Exploring Style, Object, and Environment Variation

The dataset comprises 30 distinct object-environment pairs rendered in photorealistic and 10 different styles. The styles used in the dataset are pixel art, watercolor, sketch, voxel art, charcoal drawing, embroidery, neon glow, mosaic art, graffiti and glass painting. All the images have been generated using Stable Diffusion XL (SDXL). The images generated by SDXL have dimensions of 1024x1024. These images were resized in different dimensions such as 512, 384 and 256 and also provided. For each style, multiple variations of the object environments have been generated. Experimentation was done with 3 guidance scale values (5, 8, 11) for image generation. With respect to the prompts for image generation, the prompt structure was varied. The style and style description were placed in 3 various positions in the prompt: start of the prompt (prepend), middle of the prompt (mid) and end of the prompt (append). 5 variants of images were created. Two types of control images were also created: 1. Style_only (these images don't have any object environment pair) 2. Content_only (these images don't have any style associated with it in the SDXL prompts). All the image resolution levels (1024, 512, 384, 256) have 12915 images each. All the metadata related to the images have been provided in the csv file. The 30 distinct object-environment pair images and metadata are distributed across 5 chunks of files. The v1 files contain 8 distinct object-environment pairs, v2 files contain 6 distinct object-environment pairs, v3 files contain 6 distinct object-environment pairs, v4 files contain 5 distinct object-environment pairs, and v5 files contain 5 distinct object-environment pairs. The dataset has multiple usecases, though this is not an exhaustive list: Style classification and clustering Evaluating prompt faithfulness of diffusion models Evaluating style-conditioned image generation capabilities of diffusion models Cross-style image retrieval Style transfer evaluation Prompt sensitivity analysis