Tag
DecQ introduces lightweight detail-condensing queries to improve reconstruction and generation in representation autoencoders without disrupting pretrained semantic spaces.