Take a look at these images. They are images of people right? Well, No.

These realistic looking faces were created by Nvidia, using their state of the art GANs.

Now you might be wondering what is GANs! Here, let me tell you...

GANs stand for "Generative-Adversial Networks". These are a special kind of deep learning models. They are so efficient that they can create realistic looking human faces.

Continuing, let's talk about their structure...

GANs don't work like any regular machine learning model. At it's core it has 2 units. One is called a "Generator" and the other is called a "Discriminator".

Let's see how they work.

Let's take an example of realistic face generation.

Here the generator will create a face and then it will send it to the discriminator. The discriminator will have a loss function that will check whether the generated face is realistic enough.

Let's take an example of realistic face generation.

Here the generator will create a face and then it will send it to the discriminator. The discriminator will have a loss function that will check whether the generated face is realistic enough. Then the loss function of discriminator will tell the optimizer of the generator to create more realistic faces and this cycle continues until the loss function is satisfied that generated image is realistic enough.

This is how GANs are able to create such realistic images.

This deep learning model might seem plausible but there are certain limitations to it:-

• First of all, However good this model is, it takes up a lot of resource to generate output.

Why? Read Next..

• It does a lot of computation to generate results and those computations can only be done using TPUs(Tensor Processing Units).

• Because if the above reason, meaningful results can only be given by someone with a lot of resources. Like big tech companies.

This is the end of the thread I guess. More awesome stuff coming ;) Meanwhile, why don't you follow @hrdkcodes, it helps a lot.

Stay tuned :)