Contributions

We extend the GAN model to tackle the fashion search with attribute manipulation, dubbed as AMGAN, which is able to unify the prototype image generation and the target fashion item search in an end-to-end manner.

We adapt the discriminator with the semantic discriminative learning and adversarial metric learning, which is able to not only regularize the correct attribute manipulation on the generated prototype image, but also enhance the metric learning for fashion search with attribute manipulation.
 

Extensive experiments on two real-world datasets validate the superiority of the proposed AMGAN. Furthermore, our model can be easily applied to other image search with attribute manipulation tasks. As a byproduct, we released the codes to benefit other researchers.