Seri Lee, Eun Jee Sung, Juhee Lee, Jieun Park


UGly-Net (2019) is a playful interactive tool that invites users to tinker with and create their own video synthesis U-Nets. In doing so, users can learn by changing the feature maps of the model and experiment on their creative expression using deep neural networks.

Artificial intelligence (AI) is increasingly distancing itself from the general public with its rising complexity. Although AI possesses rich potential as a creative medium, it is seen by many people as intimidating rather than inspiring.

UGly-Net introduces the U-Net structure in simplified visualization with bold strokes and vivid colors. It is intended to offer users a playful, humorous, and creative take on deep neural networks. Users can start from a well-trained and stable U-Net for slow-motion video synthesis, and then transform the network by freely making modifications to each part of the network. Users can dictate how to “carve-out” the inputs of each layer, resulting in unique network for each user. UGly-Net changes the network feature maps according to user-generated variations to synthesize videos with glitchy effects. The videos are then collected and displayed in the gallery along with its corresponding network structure.

In the process of “doughing” the U-Net iteratively and synthesizing videos, the user instinctively learns the role of each part of the network and its effects on videos. While previous digital generations manipulated retro computers and formed their own understanding of computers, our work can help users understand the new “untrodden realm” of neural networks.

UGly-Net is also an artistic means of expressing individuality. Contemporary computers deliver flawless images that are nearly uniform in quality across operating systems and hardware. At this time, the glitch created by the user is a unique creative expression that only the user’s own video and set of vectors between layers can produce.

ugly-net_additional_media_2 (1)