No Longer Irrelephant

We wanted to see if a General Adversarial Networks (GAN) could input data on top selling t-shirt designs to output graphically what is appealing to the average buyer. We also wanted to gain more insights on how the data input/output is achieved for images with and without text. We focused on the elephant niche and used a web scraper to obtain images of the 1000-2000 best seller novelty elephant T-shirts on Amazon and train the GAN network to identify patterns in the text placement and imagery that humans are visually attracted to. We used a StyleGAN2 in RunwayML to train and generate the content to be displayed on the Tshirt products. The model was pre-trained with horse images and input of 724 cleaned elephant graphics and resulted in a FID score of 105.6.

Takeaway: There is a lot of underlying ‘noise’ in images so in future work we would write a script to convert all dataset images’ backgrounds to white and remove text to improve the quality of output. Overall satisfied with the turnout but honestly I wouldn’t buy a shirt with this graphic on it even if I was really into elephants. GANs are complex and there is still much to learn. The real elephant in the room is that Runway ML is a bit shady in that they don’t provide a transparent way of informing how much money users are spending during a training session.

Deep Space; Human-Robot Interaction; Haptic Feedback; Space Exploration.