The NFT images are generated using a DeepDream CNN. This deep neural network enhances images by over-interpreting patterns in the original image, thus creating a “dream-like” image. The idea in DeepDream is to choose a layer (or layers) and maximize the "loss" in a way that the image increasingly "excites" the layers. Choosing lower layers produces simple strokes or patterns and choosing higher layers produces more abstract, convoluted patterns. This is because lower-level features are learned in earlier layers in the base model and higher-level features are learned in later layers! In our case, we chose lower-level layers. We then created a feature extraction model and calculated the loss at these lower layers (loss = sum of activations at specified layers only). Whereas in a normal CNN you aim to minimize loss (via gradient descent), in DeepDream we aim to maximize loss (via gradient ascent)! This is because we want to add perturbations and stroke patterns to the image instead of detracting noise. We ran the DeepDream model on images of famous bodybuilders (Ronnie Coleman, Arnold Schwarzenegger, Larry Wheels) and the images you see above and below are the result.