Tag
研究发现语义泄漏（semantic leakage）不仅源于概念与实体的直接归属，即便概念只是出现在场景中也会驱动模型生成刻板印象输出，这揭示了一种当前性别偏见评估所忽略的联想性偏见机制。
This paper investigates semantic leakage in audio-video diffusion models through the 'attention triangle' of cross-attention mechanisms, and presents methods to enhance semantic grounding during generation.