In the context of AI image and video generation, noise refers to random visual information - pixel-level randomness similar to the grain on film or static on an untuned television - that diffusion models use as their starting point for generation. Denoising is the core process by which these models transform that random noise into coherent, structured imagery through a series of learned refinement steps.

Diffusion models work by learning to reverse the process of adding noise to an image. During training, the model is shown images with progressively more noise added until they become pure randomness, and it learns to predict and reverse each step of this degradation. At generation time, the model starts from random noise and applies its learned denoising process iteratively, guided by the text prompt or other conditioning information, gradually transforming noise into a coherent image or video frame. The number of denoising steps taken during generation affects the quality and coherence of the output - more steps generally produce higher quality results at the cost of longer generation time. This process is why generation models of this type are called diffusion models.

The noise-to-image generation process has practical implications for creators. The starting noise state is often called a seed, and using the same seed with the same prompt and settings will produce the same or very similar output, enabling reproducibility. Varying the seed while keeping other settings constant explores different interpretations of the same prompt, which is a useful technique for finding the best version of a generation without changing the underlying creative direction.