Training teaches the model to destroy pictures in a controlled way and to undo that destruction. Generation runs the undoing on pure noise, steered by your prompt, your reference image and whatever structural conditioning you supply. That is the whole idea; everything else is engineering.
Two practical consequences follow. First, output is stochastic: the same prompt gives different frames unless you fix the seed. Second, the model has no concept of the object it is drawing, only of what pictures of it tend to look like, which is exactly why labels, spoke counts and hands need gates.
How does a diffusion model actually generate an image?
It starts from random noise and repeatedly predicts what to remove, guided by your prompt and any conditioning, until a coherent frame is left. Each pass is a denoising step.
Why does the same prompt give different images?
Because the starting noise differs each run. Fix the seed and the same prompt with the same settings reproduces the same frame.