In a previous blog, I asked the question of whether names are somehow connected with certain faces. CLIP projects text and images to the same latent space. I decided to see what happens if the names Karen and Kevin are projected into the StyleGAN2 latent space and then visualised by the generator. I must admit that the samples are slightly cherry-picked. Entering a single name is a slight abuse of the method which normally expects a little more elaborate description. Nevertheless, enjoy!
Kevin 
Also Kevin
Kevin has experienced some shit. Probably due to the self-fulfilling prophecy of Kevinism. In the previous blog on ideophones, the figure on the right was also named Kevin only second to Damien. Neither of these Kevin’s look like the Kevin that I have in mind.
I pity the manager this Karen is talking to. I also wondered what Matige Kunstintelligentie personified would look like. Note that CLIP is trained in English and that these Dutch words will unlikely be part of the vocabulary. English is very easily tokeniseable whereas Dutch has compound words like Kunstintelligentie. I wonder whether that has a significant impact on Dutch NLP.
This technique of generating faces with CLIP and StyleGAN2 from names definitely needs some polishing work. As stated before this is not the intended use of CLIP. However, with a labeled dataset it could be done. If someone is willing to (legally) scrape some database with names I will happily train the model that does this. Then we might truly find out who is most Karen or Kevin.