Applying Generative Adversarial Network (GANs) in The Fashion Industry

There are still many theoretical challenges in the design of GANs, so authors of research papers do not
focus their attention on the tasks that have an immediate practical impact. For instance, there were papers
investigating the reproducibility of reported results from different GANs, generation of clustered data with
GANs, accelerating GAN computational operations on specific hardware (Field Programmable Gate
Arrays), and many other works.

However, there have also been some interesting ideas that may be of more practical value but have not yet
made it into the business that we would like to review.

One interesting topic to cover is applying GANs in the fashion industry. The publication is dedicated to AI
technologies for virtual fitting room development. After deep research, it was decided that any virtual fitting
room application must be combined with AR libraries for additional functionality. Therefore, data scientists
should work on a model for pose estimation that could detect crucial points of the frame.

It seems that fashion companies have slowly become aware of the benefits machine learning can bring and
are looking for new ways of enhancing the customer experience with its help.

For instance, Zalando is researching how GANs can be used to generate images of fashion models
wearing customer-selected clothing items, thus helping customers visualize how the individual outfit
items would combine together. This is an excellent example of how GANs can change the industry since in
the past the only option to solve this problem would be to take photographs of a model wearing each
combination of clothing which is simply not feasible.

One additional curious piece of research from the same company shows that GANs can also be applied to
help designers experiment with new clothing ideas by allowing them to modify the shape, color, and texture
of an existing clothing item.

Continuing on, we would like to discuss a popular topic of face swapping. While almost everybody has
heard of the DeepFake technology as it has been used to make fun of politicians and celebrities, not
everybody knows about the technology’s limitations.

For each target and source subject, the DeepFake network has to be trained on video sequences for both
subjects that are not always available. In a recent paper on the topic, the authors introduced a new face
swap model called FaceShifter that allows users to swap faces in situations when only two photos are
available.

Moreover, the network is able to handle situations when the target or source face is occluded by clothing items, obstacles, or when the lighting conditions significantly differ between the two images.

One can think of negative consequences a technology like this may lead to when made accessible to a
wider public, but there are positive applications as well. For instance, it can be used to anonymize peoples’
faces in photos/videos by swapping the real faces with synthetic ones as in the research published here.
The resulting media could be used for any purpose (e.g. content creation) without being afraid of legal
claims and objections from the persons originally captured in the media.