Real and Fake AI-Generated Standardized 512px Image Forensics Dataset

HybridForensicsNet is a curated dataset of 10,000 high-resolution (512\u00d7512) images designed to benchmark digital image forensic algorithms against a hybrid threat model. The dataset is strictly balanced (50\/50) between authentic photography and AI-generated synthesis. The Real class (5,000 images) comprises high-quality portraits from FFHQ and diverse scenes from MS COCO. The Fake class (5,000 images) aggregates samples from four distinct generative architectures: ProGAN and StyleGAN3 (representing GAN-based threats) alongside Stable Diffusion XL (SDXL) and Midjourney (representing Latent Diffusion Models). By incorporating both older GAN artifacts and modern diffusion-based textures, this dataset enables the training of robust detectors capable of generalizing across evolving generative paradigms. All images are standardized to JPEG format with uniform compression and resolution to facilitate direct input into deep learning pipelines.