Fix the Noise: Disentangling Source Feature for Controllable Domain Translation

Figure 1 . Given a source model Gs, our method can smoothly control the degree of source domain features in a fine-tuned model Gs→t. Samples in each row are generated from the same latent code z ∈ Z by Gs and Gs→t. Here, Hs, Hs→t, and Ht denote feature spaces of the source, our model, and simply fine-tuned target model, respectively. Our approach explicitly guides the model to preserve the source features by using the anchor point n anch , which allows a flexible and smooth cross-domain control via Gs→t.