TruEye detects AI-generated humans with 5-category classification, 100x faster than LLMs

TruEye is a novel AI detection model from researchers Jay Barot and Dan Lin that addresses the growing problem of AI-generated images used for fraud and social engineering. Unlike conventional detectors that output a single binary label (real vs. fake), TruEye classifies images into five distinct compositional categories: fully AI-generated human, real human in AI-generated scene, AI-generated human in real scene, real human composited into a real scene where they were never present (the hardest case), and fully real. This fine-grained approach is powered by a mask-conditioned dual stream transformer that separates human and scene tokens while preserving patch-level spatial correspondence. Specialized reasoning within each stream and region-gated cross attention enforce semantic coherence between subject and background, while token-level supervision and global compositional classification yield interpretable predictions without invoking large language models.

On a technical level, TruEye’s key innovation is restricting intra-stream attention to semantically coherent tokens, which dramatically speeds up inference—over 100× faster than competitors that rely on LLMs for explainability. The model was evaluated on six existing datasets plus a newly curated FineSyn dataset, and it surpasses state-of-the-art detectors in accuracy, inference speed, and generalization to unseen AI-generated or manipulated images. The paper, published on arXiv (2606.27505), is 18 pages with 3 figures and includes code and data links. By enabling precise localization and categorization of synthetic human content, TruEye offers a practical tool for platforms fighting misinformation, fraud, and deepfake-based identity theft without the computational overhead of LLM pipelines.