If we're talking about directly using the current text to image generators, thats true. But that's not what I'm talking about.

The real fun starts when you start connecting them to language models

https://twitter.com/jaukia/status/1567914039061217287

This is an example with GPT3. But will get better in the coming years ( year? ) with even more powerful LLMs, especially the multimodal ones that can directly inagine visuals instead of just text that is passed to the text-to-image renderer ( the rendering model might still be used for the final composition but it helps to have the designer model be able to think in actual visual terms ).