Uncanny Rd. is a drawing tool that allows users to interactively synthesise street images with the help of Generative Adversarial Networks (GANs). The project was created as a collaboration between Anastasis Germanidis and Cristobal Valenzuela to explore new kinds of human-machine collaboration that deep learning can enable.
The project and uses two AI research papers published last year as a starting point (Image-to-Image Translation Using Conditional Adversarial Networks by Isola et al. and High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs by Wang et al.) Users are asked to interact with a semantic colormap of a scene, where each color represents a different kind of object label (e.g. road, building, vegetation, etc.). The neural network model was trained using adversarial learning on the Cityscapes dataset, which contains street images from a number of German cities.
There are advantages and disadvantages to this new paradigm that are apparent to anyone who plays with Uncanny Rd for a few minutes: on the one hand, the neural network can produce images of astonishing fidelity from a very generic high-level representation (the semantic map). On the other hand, the user has very limited control over how the final image will look like — which makes the interface useful for brainstorming and exploration but not for bringing to life a specific creative vision. The way forward is finding ways to combine the power of neural networks with more traditional symbolic approaches.
The results are poetic and surreal composites of uncanny locations, that appear to retain a sense of place (thanks to source material) regardless of their immateriality.