What Makes Pixel Planets So Special?

Pixel Planets are not an ordinary NFT. They are the culmination of bleeding edge technology in the realm of machine learning and AI. Pixel planets are your very own unique piece of the galaxy.

Before we begin, we need to understand some machine learning and AI concepts. Lets run through the process behind the hood.

When it comes to image generation, the best tool by far is a Generative Adversarial Network (GAN). GANs are a type machine learning algorithm that use neural networks to generate “fake” images that resemble the training images in a believable fashion. There are two main parts to a GAN, the discriminator and the generator. The discriminator is a neural network which will compare an authentic source image to a fake image and guess whether the fake image is truly fake, or if it is close enough to the original to be believable. The generator is another neural network which is responsible for generating these fake images. It accomplishes this by mixing random noise within the neural network to hopefully generate a believable fake image. Initially, the generator produces very bad fake images and the discriminator is able to tell that it is not an authentic image with ease. Over time however, these incorrect results are fed back into the generator through a process called back propagation where it is then able to learn from its mistakes and produce more authentic fakes. If all goes well, the generator will improve to the point where the discriminator is not able to tell that the fake image is fake anymore. Once this occurs, we cease training the model and we are then able to produce unique images which resemble the theme of the original sample images.

Now that we understand GANs, let’s see how they are used to generate Pixel Planets. GANs are utilized twice within the Pixel Planet pipeline, and they are slightly more complex variations of the GANs we went over earlier. The first GAN is a DCGAN (Deep Convolution GAN) which uses a training dataset for planet images and generates the authentic fake images. We then take these fakes and run a them through a CycleGAN. A CycleGAN is able to take a source image and transform its theme and style into something completely different. For Pixel Planets, we used the CycleGAN to transform the outputs from the DCGAN into pixel art. One caveat of this transformation is that the pixel art is not exactly pixel art, the level of detail is too high for what we want. So as one final step we run the image through a custom algorithm to transform it into pixel art, it is at this step that we introduce our signature color palette.

But ML isn’t perfect, so we get some really really ugly ones too. We spent many hours combing through and curating the best of the best Pixel Planets that you’ll be proud to own one day.

So all in all, we spent 1450+ combined hours training the ML models and tuning the hyperparameters, as well as way too many hours manually sifting through the thousands of generated images to select the best Pixel Planets.

Related posts:

Chimneys are an essential part of many homes, providing warmth and comfort during the colder months. However, over time, your chimney may require repairs to ensure it remains safe and functional. In…

Children deserve our excellence. In my journey of motherhood, I am often inspired by mothers who tell me “I wish I knew this…” or “I wish I did this…”, implying that with knowledge and experience…

Best Travel Agents Egypt is a country that tops the bucket-list country for many travelers. It’s a country stuffy with unreliable relics of ancient history, from the Pyramids and Great Sphinx of…