Good engineering. Midjourney likely has a lot going on under the hood before your prompt actually gets to Stable Diffusion. As an example you can check out this

Good engineering. Midjourney likely has a lot going on under the hood before your prompt actually gets to Stable Diffusion. As an example you can check out this research paper [0] which seeks to add prompt chaining to GPT-3 so you can "correct" it's outputs before it reaches back to the user. There's also no rule that states you can only make one call to SD, MJ likely bounces around a picture through a pipeline they've tuned to ensure your generated image looks more reasonable. [0]: https://arxiv.org/abs/2110.01691 https://arxiv.org/abs/2110.01691