This is an embarrassingly idiotic article. They're comparing against the least intensive text model, ie some million parameter model nobody uses. Stable Diffusion is a 3.5 billion parameter model, while the GPT models a billion people are using for text generation are over 10 trillion parameters. To say nothing of the differences in average context size usage. The actual ratio of SDXL to text generation pollution is probably literally reversed from what the article claims by lying with statistics.

(Note, however, that OpenAI's image model is much much larger than SDXL; however, we don't have precise numbers for it. Nonetheless, misinformation is misinformation.)