Unified Detection of Synthetic and Manipulated Images via Dual-Stream Artifact Fusion

Generative image models and modern editing tools make it increasingly difficult to verify the authenticity of visual content, especially when images span fully generated, locally edited, and hybrid, partly synthesized cases. To address this unified setting, we construct NESS, a large-scale benchmark that aggregates public datasets into four categories: Natural, locally Edited, fully Synthesized, and locally Synthesized, resulting in 210,776 images and a balanced evaluation protocol. On top of NESS, we introduce DAFNet, a dual-stream artifact-fusion network that couples an RGB semantic stream with an artifact stream built from fixed SRM high-pass residuals, and integrates them via a lightweight spatial gating module to adaptively exploit complementary cues across manipulation types. On the NESS test set, DAFNet achieves a balanced accuracy of 0.9191 and a ROC-AUC of 0.9799, outperforming the evaluated baselines under the same protocol. The NESS dataset and DAFNet implementation will be released to support future research on robust image authenticity assessment.