Out-of-distribution (OOD) detection is necessary in various real-world applications. However, training artificial intelligence models exclusively on in-distribution (ID) data frequently results in misclassifying OOD samples as ID classes, leading to significant adverse consequences. Inspired by recent studies on spurious outlier generation, we propose two novel strategies for generating fake OOD data. The first strategy constructs fake OOD data from the perspective of foreground-background separation by a large vision language model, while the second strategy combines patches from different ID images to create a new composite image as fake OOD data. To further improve the model’s awareness of OOD data, we design a novel loss function and a novel scoring function from two different perspectives to separate ID from OOD data. Our method demonstrates superior OOD detection performance on three widely used benchmarks.