Product developers nowadays are able to automate product and service design in a large scale without consumer preference data. However, such design may not meet the needs of heterogeneous consumers with dynamically evolving preferences. Leveraging a unique unmanned photo gallery setting where consumers choose templates to self-take photos, we study how the design templates can be automated efficiently in large scale, while incorporating heterogeneous consumer preference. Our approach can improve the product design when facing “cold-start” problem by leveraging external data of user generated content (UGC) from external websites. We first estimate a “deep” choice model using deep neural network, and then we collect a set of external photos and remove human subjects to create templates that are similar to the internal templates offered by the firm. We then leverage both internal and external data to generate new templates. We compare templates generated using generative adversarial network (GAN) without incorporating consumer preferences and advanced generative models, conditional-DCGAN, with preference information. We show how incorporating preference information can improve the quality of generated templates.