UniGenDet: A Unified Generation-Discrimination Framework for the Co-evolution of Image Generation and Generated Image Detection

UniGenDet incorporates image generation and generated image detection into a unified generation-discrimination framework, enabling the two to co-evolve and mutually reinforce each other. The framework designs a symbiotic multimodal self-attention mechanism and a unified fine-tuning algorithm, and introduces a detector-information-guided generation alignment mechanism to enable information exchange between the two tasks. Experiments on multiple datasets show that it achieves SOTA performance.