Synthetic Image Physical Properties Detection (SIPD) dataset

The SIPD (Synthetic Image Physical Attribute Detection) dataset is designed to provide a structured, controllable, and physically annotated image resource for research on synthetic image detection. The dataset contains 22,203 micro-landscape images, including 15,012 real captured images and 7,191 manually synthesized images, with resolutions of 8256×5504, 3656×2664, and 2560×1440. Real images are captured using a Nikon Z7 camera paired with a Tamron 90 mm f/2.8 macro lens to ensure high-resolution and optically realistic imagery, while synthetic images are manually created in Photoshop to achieve precise control over editing operations and physical attributes.The dataset is divided into Street View and Portrait groups. The Street View group covers a wide range of vehicle types and colors, with paired replacement sets that support realistic substitutions during compositing. The Portrait group includes images captured against black, sky, object, and real architectural backgrounds, using color-diverse miniature figures to present complex foreground–background relationships. Both groups provide detailed annotations of essential controllable physical attributes, including light source position, color temperature, camera position, object distance, depth of field, and background complexity, enabling researchers to investigate model behavior under individual or combined variations of physical parameters.The synthetic portion of the dataset includes three representative editing methods—splicing, copy-paste, and erasing—and follows a single-variable principle to generate images for six categories of physical attributes. A total of 1,671 synthetic images are created for the Street View group and 5,520 for the Portrait group, with balanced distributions across editing methods and attribute categories to support systematic comparative studies.All images are partitioned into training, validation, and test sets using a 7:2:1 ratio, ensuring sufficient training volume and consistent evaluation benchmarks. As a physically controllable synthetic-image detection dataset, SIPD offers a high-quality, finely annotated resource for research on authenticity assessment, physical consistency analysis, and tampering detection in complex visual scenes.