Answers are grounded in this paper and cite their sources.
Generative adversarial networks (GANs) and diffusion models can now produce images and videos that are difficult to distinguish from real recordings. This has consequences for journalism, legal evidence, and personal privacy. Many detectors based on convolutional neural networks (CNNs) and transformers have been proposed, but three problems persist. Detectors trained on one family of forgeries often fail on unseen GANs or diffusion models. Detectors that use only pixels ignore the spectral traces that synthesis pipelines leave. Image and video forgeries are usually handled by separate models. This paper presents FreqFusion, a detector designed to address these three problems. A frequency branch based on the discrete cosine transform (DCT) and the discrete wavelet transform (DWT) runs in parallel with a semantic branch built on a pretrained Vision Transformer (ViT). The two branches are combined by a symmetric cross-attention module, in which each branch attends to the other, instead of being concatenated at the end of the network. For video, a small temporal transformer is trained on the per-frame fused tokens while the spatial encoders remain frozen. FreqFusion is evaluated on FaceForensics++, Celeb-DF (v2), DFDC, DeeperForensics−1.0, and DiffFaces, a set of diffusion-edited faces assembled for this study. The average cross-dataset AUC is 92.7%, 8.1 points above the strongest retrained baseline, and the AUC on unseen diffusion forgeries is 88.4%, 7.4 points above the same baseline. In the ablation, the fused model outperforms either branch alone, and cross-attention fusion outperforms late concatenation of pooled features. These two fusion variants also differ in pooling and in the number of parameters. The gain is attributed to fusion before pooling as a whole, not to the cross-attention operator alone. Accuracy is somewhat reduced by JPEG recompression, blur, and added noise. A single trained spatial backbone serves both image and video detection. The experiment scripts, the recipe for building the diffusion test set, the reference implementation and its unit tests are publicly available.
No takes yet. Share an insight, caveat, or question.
Yadav et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: