TF-IDF not a very effective metric for small corpus. Most words end up having the same value.
Abstractive Summarization vs Extractive summarization? We need a better method to find important words in a sentence.
Retrieving an icon, taking context into account. (Homonyms (play/outgoing/light), not ..)
Is layout really important? If coherence is necessary, then no-overlap is a must.
How can we use GANs in this case to compare different arrangements?
GANs seem to be most useful at generating data, given a small amount of tattoos, can we use it to generate more tattoo like images (given the lack of tattoo datasets)?