Threaded History (2020). We travel through a carefully-crafted path in the latent space of a model trained to generate the details of huipils from Guatemala and Mexico, revealing a beautiful history embedded in the threads. In it, we visit both specific colors, as well as specific huipil patterns in order to fully explore the capabilities of the trained model. Moreover, patterns from specific regions can be recognized lying dormant in-between different known huipils and colors, showing us a hidden history of shared cultures, perhaps also of those that may yet come to pass. This work seeks to explore the former, revealing the history that lies within, and hopes to open the question of what more can we learn from images alone.
All frames in the videos were generated using the official TensorFlow implementation of StyleGAN2 by training it with the details of huipils from Guatemala and Mexico. Previous work had focused on generating the huipil as a whole, but here we explore the minute details, the threads that compose each huipil. Indeed, while generating the whole huipil, these details were lost and became part of the noise, effectively erasing or covering the history threaded within. More specifically, in this work we traverse the disentangled latent space W, therein interpolating smoothly and spherically. This way, we can better showcase the capabilities of the model, as well as avoid confusing the observer with which images we wished them to focus on. This movement throughout W is precisely controlled to generate what we wish to portray, as well as how quickly to change between each different disentangled vector.
The trained model, as well as the dataset used to train it, will not be released due to ethical reasons: after all, these traditional garments are not mine to give away, merely to showcase their beauty, especially since they contain the history, knowledge, and legends of numerous cultures.