The use of GANs, an AI model in which two parts reinforce each other to create a better original, has been taken to a higher level by Amazon for use in e-commerce. The so-called ReStGAN could help buyers of products find better.
The AI generates a picture of the product being searched for by means of a search query, after which the customer provides a better picture of what is being searched for by adding additional search terms. The example given by Amazon is the search term ‘black trousers for women’. By subsequently adding terms such as ‘small’, the image is adjusted to include these criteria as well.
Amazon placed the system alongside other similar text-to-picture systems and believes that the end result is significantly better: according to Amazon, the application of colours in particular went smoother with the ReStGAN.
The two-part element of the model forms the basis of the improved results. By means of a GAN, a standard model is first created on the basis of the search term in a low resolution. Then that model is lifted to a higher resolution with better colors by another GAN. In order to make the process smoother, the system is defined per part: pants, jeans and shorts.