FootwearGAN experiments
experiments | 2 Days | May 2020
A prompt to design 'futuristic footwear', from my friend @Prabhash Dhama, led me to explore Machine Learning and how it can help designers. These are my first experiments with Machine Learning and GAN(Generative adversarial networks).
Dataset
In order to generate new images, the machine needs to be fed with sample/dataset. I've used 2,700 images of sneakers from 98mprice's Project (linked here). These images have been sourced from online shopping platforms, they are mainly top brands from Adidas, Nike, Reebok, Vans, Asics and many more.
Process
Usually for such tasks people use Keras(with tensorflow backend) and I got similar results as 98mprice's Project. I tried RunwayML's FaceGAN (learn more about Runway ML here) for the process and their trained model can produce fabulous results without coding a single line. Later that month I discovered Zixiong Wei's FootwearGAN on Runway and it creates amazing results as well.
Usually for such tasks people use Keras(with tensorflow backend) and I got similar results as 98mprice's Project. I tried RunwayML's FaceGAN (learn more about Runway ML here) for the process and their trained model can produce fabulous results without coding a single line. Later that month I discovered Zixiong Wei's FootwearGAN on Runway and it creates amazing results as well.
Results
Below are 100 images generated result of 4 hours of computation and 3000 steps of processing.
Conclusion:
I see this technology as a mere tool to facilitate designers and creators in their process. And not a replacement of the people its meant to be used by. As we can see the machine can come up with 'new' images but nothing drastically 'new' because of the model its trained on. For more radically new images/styles to produce, it needs to fed with exquisite dataset. My next experiments will to be explore what the machine can produce from such data.
