Recent advances in diffusion models have enabled the generation of highly realistic images, posing greater challenges to developing generalizable detection methods for identifying images from unseen diffusion models. Current research has identified two types of artifacts for synthetic image detection: fingerprints, which are specific to the model that generates the image, and universal artifacts, which are shared across different generative models. However, existing detection schemes predominantly rely on fingerprints to distinguish synthetic images, resulting in limited generalizability, as universal artifacts are particularly critical for enabling generalizable detection. To address this limitation, we propose Interpolated Diffusion Image Reconstruction (IDIR), a novel approach for extracting universal artifacts. IDIR reconstructs an image using a diffusion model, allowing the differences between the original and reconstructed images to be regarded as artifacts. By narrowing these differences through interpolation, the fingerprints and universal artifacts diminish proportionally, and their amplitude gap attenuates. This step reduces the dominance of fingerprints while enhancing the relative prominence of universal artifacts, mitigating detectors’ bias toward fingerprints. We trained a lightweight classifier on IDIR-processed images to capture universal artifacts. Tests on the latest benchmark show that our classifier outperforms the best baseline by 11.55% in average accuracy.