FALCON: Forensic-Aware Language-guided Contrastive Learning for Generalized Synthetic Image Detection

Quang Huy Nguyen$^1$, Keun Ho Ryu$^2$, Mutia Delina$^4$, and Van Huy Pham$^4$

$^1$Natural Language Processing & Knowledge Discovery Laboratory, Faculty of Information Technology, Ton Duc Thang University, Ho Chi Minh city, Vietnam
$^2$Data Science Laboratory, Faculty of Information Technology, Ton Duc Thang University, Ho Chi Minh city, Vietnam
$^3$Physics Department, Faculty of Mathematics and Natural Sciences, Universitas Negeri Jakarta, Jakarta, Indonesia
$^4$School of Technology, University of Management and Technology Ho Chi Minh City, Ho Chi Minh city, Vietnam

---

News

• [Sep 9, 2026] 🎉 Our paper has been accepted to FITAT 2026 (The 18th International Conference on Frontiers of Information Technology, Applications and Tools), held in Cheongju, South Korea (October 30 - November 2, 2026).

Abstract

Recent advances in generative models have significantly improved the realism of synthetic images, making cross-generator synthetic image detection increasingly challenging. Existing forensic detectors often learn generator-specific artifacts and therefore suffer substantial performance degradation when evaluated on unseen architectures. In this paper, we propose FALCON (Forensic-Aware Language-guided CONtrastive Learning), a language-guided forensic framework that uses forensic-aware textual supervision to learn transferable visual representations. FALCON aligns image features with textual descriptions of semantic content and forensic traces, and then refines the detector with a hybrid contrastive and classification objective. To evaluate cross-generator robustness, we construct UniRF-112K, a benchmark of 112,000 real and synthetic images spanning GANs, diffusion models, transformer-based generation, and flow matching. Under a one-generator training protocol, models are trained on ProGAN and evaluated across diverse held-out generators. Experimental results show that FALCON achieves the highest mean AP of 83.30% among the evaluated methods. The dataset and supplementary resources are publicly available at https://github.com/VictorNguyenLPN/FALCON.

Dataset

• Name: UniRF-112K
• Scale: 112,000 images collected from 9 state-of-the-art generators
• Source: https://www.kaggle.com/datasets/daddychillonkaggle/unirf-112k
• Protocol note: Zero-shot setting (train on ProGAN, test on unseen architectures)

Citation

If you find this work useful, please cite:

@misc{nguyen2026falcon,
  title  = {FALCON: Forensic-Aware Language-guided Contrastive Learning for Generalized Synthetic Image Detection},
  author = {Nguyen, Quang Huy and Ryu, Keun Ho and Delina, Mutia and Pham, Van Huy},
  year   = {2026}
}

Corresponding

• Van Huy Pham: [email protected]

Or you can create an issue in this Github Repository