I have played around with GPT quite a bit and I would say that GPT understands the difference. Text-to-image models are not specialized in the text-parsing part, so I think it's forgivable that they are not as good at it.
Edit: Actually I tried this right now with two prompts, and I was wrong. It might still be that gpt understands compositionality but the prior that people ride horses is just that strong. But what I saw was that with this particular situation the model got it wrong.
Edit 2: With some heavy hinting it managed to understand the situation. Italics mine. "An astronaut is walking on all four. A very small horse is sitting on top of him, riding him even. Shortly after the astronaut stops, exhausted.
The horse is too heavy for the astronaut to carry and he quickly becomes exhausted. Next, the horse gets off the astronaut, stands on its own four legs, and walks away."