Could that be something to do with the things I keep reading about how somehow knowledge from, say, an LLM for generative text somehow carries over (in some way) to an LLM for image generation? I'm obviously not very knowledgeable in this area :).