I can imagine labs competing against a Pangram benchmark to prove their models can output realistic images