1 paper · 1 filter
Jiaqi Zhu, Xincheng Chen, Yuncheng Wu +2
Generative models trained on a source domain often produce samples that are poorly aligned with shifted target domains, limiting their effectiveness for target-domain data augmenta…