Improving prompt understanding in Stable Diffusion (e.g. LLM integration)

[CalculonPrime]

Even the latest SD version (SDXL I assume?) doesn't seem capable of grasping separate objects and properly assigning their attributes, and seems to lack positional knowledge derived from your prompt.

However, most of us have used ChatGPT and know that this should be a solvable problem. I see from Cornell Paper: LLM-grounded Diffusion that research is already underway towards this goal.

Let's have this thread be a place for a high-level discussion of such potential improvements to SD. Are the authors of SD already attempting a similar integration with LLM software?