Title: Venus-DeFakerOne: Unified Fake Image Detection & Localization

Abstract:

The rapid advancement of generative artificial intelligence has fundamentally altered the landscape of image forgery, erasing the historical distinctions between document editing, natural image manipulation, DeepFake creation, and comprehensive AIGC synthesis. While forgery generation has moved toward a unified approach, research into Fake Image Detection and Localization (FIDL) remains siloed and fragmented. This disconnect highlights a critical mismatch: the detection methodologies are still domain-specific, whereas the forgery mechanisms are becoming increasingly unified. Resolving this discrepancy presents two primary obstacles for FIDL: comprehending the transfer and interference of artifacts across different domains, and constructing a high-capacity unified foundation model capable of performing joint detection and localization.

To overcome these hurdles, we introduce DeFakerOne, a data-driven, unified FIDL foundation model that integrates InternVL2 with SAM2. This architecture facilitates the simultaneous identification of forgeries at the image level and the precise localization of manipulated pixels at the granular level across a wide variety of contexts. Our extensive empirical evaluations confirm that DeFakerOne delivers state-of-the-art results, surpassing existing baselines on 39 distinct forgery detection benchmarks and 9 localization benchmarks. Additionally, the model demonstrates exceptional robustness against real-world perturbations and advanced generative models, including GPT-Image-2. Finally, we offer a comprehensive analysis of data scaling laws, patterns of cross-domain artifact transfer and interference, the importance of fine-grained supervision, and the preservation of original resolution artifacts, thereby outlining the core design principles necessary for building scalable, robust, and unified FIDL systems.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC