ConvNeXt–ViT Cross-Attention Fusion: Bridging Local Texture and Global Semantics for AI-Generated Image Detection
Year: 2026
Type: conference-paper
Abstract: The growing realism of GAN- and diffusion-based image generators has made the automatic detection of AI-generated images a pressing problem in multimedia forensics. Prior work has shown that convolutional neural networks (CNNs) are well suited to capturing local pixel-level artifacts and high-frequency texture anomalies. However, Vision Transformers (ViTs) are better at simulating long-range semantic structure and global context. Combining these two strengths through dynamic, bidirectional interaction—rather than static feature concatenation—has, however, received comparatively little attention. In order to detect AI-generated images on the Stable Diffusion v1.5 subset of the GenImage benchmark, this paper suggests a Bidirectional Cross-Attention CNN–Transformer Hybrid (ConvNeXt–ViT) architecture. In the proposed model, local (ConvNeXt) and global (ViT) token representations are projected into a shared embedding space and mutually enriched through a bidirectional cross-attention mechanism before classification. Five widely used architectures—ResNet50, EfficientNet-B4, ConvNeXt-Tiny, ViT-Tiny, and MobileNetV3-Large—were trained from scratch on a balanced 24,000-image subset and evaluated as baselines under an identical protocol. The proposed model reached a test accuracy of 95.85 percent, an F1-score of 0.9585, and a ROC-AUC of 0.9914, outperforming the strongest baseline (ResNet50) by 3.25 accuracy points and the remaining baselines by 9.0–9.6 points. Efficiency analysis further showed that, despite carrying two backbones, the proposed model achieved a batch-processing throughput nearly identical to that of ConvNeXt-Tiny alone, indicating that the observed gains stemmed primarily from the architectural efficiency of the fusion mechanism rather than from increased model capacity. These findings suggest that bidirectional cross-attention fusion offers a favorable accuracy– efficiency trade-off for AI-generated image detection.
Institution: Samsun University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7224869743