THE NON-EXISTENT CHAIR SERIES: EVALUATING GENERATIVE DESIGN OUTCOMES
Wu, Jiaying Ally
MetadataShow full item record
Generative Design has been a popular topic in the design world for a while, earlier inventions like shape grammar and space syntax generate geometrical designs with sets of rules defined by the user. The latest invention of generative design is artificial neural networks like GAN (Generative Adversarial Network), which created a new logic of generative design. Earlier inventions focused on geometrical exploration with applied rules; therefore, the generated designs are calculated results. GANs, on the other hand, because of the nature of deep learning networks - are like a black box. Since there is no way of supervising what happens within, there are levels of randomness and uncertainty. GANs are also trained with images instead of geometrical shapes or forms. Making it capable of exploring colors, image depth, as well as overall composition. In a way, it changed the logical decision-making process in design into something more spontaneous. AI also enabled a new production journey map from ideation to manufacture, introducing new design opportunities. However, when it comes to evaluating generative design, most of current work are done by developers. Which focused on statistical evaluations to calculate the similarities between the dataset and the generated images. While they are valuable for improving algorithm efficiency, it may not apply to the designs. Current evaluation methods lack empathy, especially when it comes to judging and critiquing good vs. bad design. This work aims to explore the usability and applicability of generative networks by coming up with non-statistical measurable features. This work aims to answer how realistic the generated designs need to be for them to be “viable”, and for designers to be able to recognize the object for what it is. And how the pursuit of photorealism in image generation networks may not apply to the field of design.