TBH I am not familiar with today's Midjourney, but I am thinking of things like face restoration, tiled VAEs for better high res generation, and self attention.

But many tools (like controlnet, or using images as input) require more manual control.