Percept-Lens Evaluation Data - DrawBench

Percept-Lens-DrawBench Dataset Description Purpose. The Percept-Lens-DrawBench dataset is curated as part of the Percept-Lens benchmark to evaluate the generalization capability of AI-generated image detectors under prompt-induced distribution shifts. It serves as an out-of-distribution (OOD) test set, designed to probe model robustness against semantically rich prompts not seen during training. Nature. This dataset contains only AI-generated images created from text prompts sourced from the DrawBench Dataset. These prompts span abstract concepts, visual reasoning, compositional scenes, and imaginative queries. Images were generated using diverse diffusion models, including Stable Diffusion variants, and FLUX.variants, ensuring stylistic variation. No real images are included—each image is the synthetic output of a generative model responding to a high-level prompt. Scope. The dataset includes approximately 3,200 images, with each image linked to its original prompt. It is used exclusively for evaluation within the Percept-Lens benchmark and is not part of the training data. This subset is particularly useful for testing semantic generalization, where linguistic complexity drives visual synthesis beyond familiar domains.