- The colors you see are generated by ML models, not us. If you don't like them, just hit the dislike button and our ML models will soon stop creating them.
- Icon selection based on the description is not an easy problem. You have to understand the semantics of what you're trying to say and match those with an icon's appearance. We're using word embeddings and auto encoders here.
- Initially, a genetic algorithm generates a bunch of stuff but that gets pruned by ML systems. As we get more feedback, that ranking/pruning is going to be more effective.
- Which fonts to combine together is not trivial. Convolutional neural networks come to the rescue here.
- Finally, one of the additional features we are working on is going to generate icons from scratch using GANs.
I hope this answers your question. Let me know if you wanna talk more about how we go about doing everything. Always happy to chat.