Image generation works by starting the a blank input image, just noise. In ComfyUI this is called the empty latent image. In more advanced renderings you can start with a given image to edit, but this is different than that. What 'AI' does is it takes your input: "Draw me an image of a cat in a hat." and it starts at that base 100% noise image and passes through that image many times to generate what you prompted for. Each pass through the image removes some noise until a final pass is made and an image is produced. That's called diffusion. Well, apply that to text and using the power of GPUs, instead of linear growth per token inception labs is created a model, Mercury, which diffuses from noise. Then does linear passes at the end to check for obvious error, like out of placed curly brackets in code, grammar errors or other easy things like that.