UniGenDet: Unified generative-discriminative framework for co-evolutionary image generation and detection

Tsinghua-IVG proposes UniGenDet, a unified framework that jointly optimizes image generation and generated image detection, achieving SOTA on multiple datasets. The method uses a symbiotic multimodal self-attention mechanism and a unified fine-tuning algorithm to bridge architectural gaps between generative and discriminative networks. A detector-informed generative alignment mechanism enables information exchange, improving both generation fidelity and detection interpretability. Code is open-sourced.

• #image-generation
• #deepfake-detection
• #unified-framework
• #tsinghua