Virtual Interior Design Companion-Harnessing the Power of GANs

It is always a costly affair to prepare suitable interior design. Automatic interior design generation with suitable interiors will have an enormous effect on the trillion-dollar real-estate/construction industry. The impact of deep neural networks has made a leap forward where state-of-the-art algorithms utilize Generative Adversarial Networks (GANs). AI-driven innovation with GANs has many applications in creative industries such as design, be it architectural design, landscape design or interior design, the possibilities are endless. Such generated designs have the potential to drive rapid growth and profits in the design industry. The goal of this work was to generate realistic new interior room designs by training a GAN network on the IKEA Interior Design Dataset. This work faced challenges in distinguishing a good design from a poor design along with training the GAN. The goal was to overcome these challenges by incorporating human feedback and using various evaluation methods to ensure the generated designs are of high quality. The project has the potential to provide valuable solutions for the interior design industry, by generating new designs quickly and efficiently. DCGAN generated satisfactory images for 21800 epochs with FID score of 237 but had mode collapse and vanishing gradient issues. While Style3GAN outperformed by generating high resolution images for 2500 epochs with FID score of 105.