177d68f4adef163b7b123b5c5adb3c60-Supplemental-Conference.pdf

Following [2], we extend our method into multi-domain translation on AFHQ dataset, where the source domain includesCat and Wild and the target domain isDog. In this setting, similar to two-domain unpaired I2I, the EGSDE also employs an energy function pretrained on both the source andtargetdomains toguidetheinference process ofapretrained SDE.Theonlydifference is the domain-specific feature extractorEs(,) involved in the energy function is the all but the last layer of a three-class classifier rather than two-class. All experiments are repeated 5 times to eliminate randomness. The quantitative results are reported in Table 11.