DeepFake-BMA: Photorealistic AI-Generated Faces for Binary Detection and Multiclass Attribution

This dataset contains photorealistic, high-resolution AI-generated facial portraits created by the authors to support research in synthetic media forensics, deepfake detection, and generative source attribution. Existing synthetic face benchmarks focus heavily on legacy Generative Adversarial Networks (GANs) or early latent diffusion models. This dataset addresses modern generative frontiers by incorporating modern flow-matching diffusion transformers, multimodal diffusion architectures, and distilled step-adversarial models alongside GAN baselines. The repository provides distinct synthetic face portraits for each of the following 7 generative engines : - Stable Diffusion 3.5 Large (Multimodal Diffusion Transformer – MMDiT) - Flux.1 Schnell [FP8] (12B Rectified Flow Transformer) - Flux.2 Dev (Flow-Matching Diffusion Transformer) - Flux 1.1 Pro (Commercial High-Fidelity Flow-Matching Pipeline) - StyleGAN (StyleGAN2/3) (Generative Adversarial Network) - Stable Diffusion v1.5 (Latent Diffusion Model – UNet) - Z-Image-Turbo (Distilled Step-Adversarial Diffusion) All images are standardized at native $1024 \times 1024$ resolution in 8-bit RGB JPEG format (Quality Factor $Q = 95$).