During training, the conditioning signal (such as a text prompt) is randomly dropped a fraction of the time, so one network learns both conditional and unconditional denoising. At sampling time both predictions are computed, and the sampler moves further in the direction that distinguishes them, amplifying whatever makes the output match the prompt.
The guidance scale is the main quality dial users touch in tools like Stable Diffusion (default around 7.5). Higher values buy prompt fidelity at the cost of diversity, pulling samples toward archetypal, sometimes over-saturated renderings.