Research & Papers

GuangJian's DeFakerOne unifies fake image detection across all forgery types

New model beats 48 benchmarks, detects and localizes fakes from GPT-Image-2 and beyond.

Deep Dive

The rapid evolution of generative AI has created a paradoxical gap: while forgery techniques across document editing, natural image manipulation, DeepFakes, and full-AIGC synthesis are converging, existing fake image detection and localization (FIDL) research remains domain-specific and fragmented. The GuangJian Team proposes DeFakerOne to solve this mismatch. The model is a data-centric, unified FIDL foundation that integrates InternVL2 (vision-language backbone) with SAM2 (segmentation) to perform simultaneous image-level detection and pixel-level localization across all forgery types. This design allows it to understand and exploit cross-domain artifact patterns rather than being limited to a single manipulation class.

DeFakerOne demonstrates state-of-the-art performance on a massive scale: 39 forgery detection benchmarks and 9 localization benchmarks, outperforming prior baselines. Notably, it shows superior robustness against real-world perturbations and modern generators like GPT-Image-2. The paper also systematically analyzes data scaling laws, cross-domain artifact transfer and interference, the necessity of fine-grained supervision, and the importance of preserving original resolution artifacts. These insights provide design principles for building scalable, robust, and truly unified FIDL systems as generative AI continues to evolve.

Key Points
  • Integrates InternVL2 and SAM2 for simultaneous fake detection and pixel-level localization across all forgery domains.
  • Achieves state-of-the-art on 39 detection benchmarks and 9 localization benchmarks.
  • Robust against GPT-Image-2 and real-world perturbations, with analysis of data scaling and cross-domain artifact transfer.

Why It Matters

Unified detection across all image forgery types is critical as AI-generated content becomes indistinguishable from real.