The paper introduces FakeInversion, a method for detecting synthetic images generated by unseen text-to-image models using features extracted from a pre-trained Stable Diffusion model. The method leverages an inverted latent noise map and a reconstructed input image as additional input signals. FakeInversion is trained on fake images generated by Stable Diffusion and real LAION images, achieving state-of-the-art performance in detecting high-fidelity text-to-image generators. The authors also propose SynRIS, a new evaluation protocol that uses reverse image search to ensure that the detector is not biased towards specific themes or styles. This protocol is more reliable in evaluating detectors' performance on closed-source text-to-image models. The paper demonstrates that FakeInversion outperforms existing methods on both academic benchmarks and the new SynRIS-based evaluation, highlighting the importance of inversion features for generalization.The paper introduces FakeInversion, a method for detecting synthetic images generated by unseen text-to-image models using features extracted from a pre-trained Stable Diffusion model. The method leverages an inverted latent noise map and a reconstructed input image as additional input signals. FakeInversion is trained on fake images generated by Stable Diffusion and real LAION images, achieving state-of-the-art performance in detecting high-fidelity text-to-image generators. The authors also propose SynRIS, a new evaluation protocol that uses reverse image search to ensure that the detector is not biased towards specific themes or styles. This protocol is more reliable in evaluating detectors' performance on closed-source text-to-image models. The paper demonstrates that FakeInversion outperforms existing methods on both academic benchmarks and the new SynRIS-based evaluation, highlighting the importance of inversion features for generalization.