Quantitatively (we used CLIP embeddings+cosine similarity) GPT does better than the average artist (but not top artists!), but since we used a single instruction prompt its strategy is usually the same: make everything "vibrant" and underwater 😁
What does all this mean?
