In the context of AI image and video generation, particularly diffusion models, noise refers to random statistical variation added to or present in image data: mathematically similar to the visual static of an untuned analogue television or the grain of underexposed film. In diffusion models, noise plays a specific and central role in the generation process: the model is trained to reverse a noise-addition process, learning to progressively remove noise from a noisy image until a clean, recognisable image emerges. Generation begins with pure noise: a tensor of random values with the same dimensions as the target image: and the model iteratively predicts and removes components of this noise over a series of steps, guided by the text prompt or other conditioning signal, until a structured, coherent image is produced from what began as complete randomness. The noise level at any given point in this denoising process represents how much noise remains, with the initial step being maximum noise and the final step being the fully resolved image.
Noise level becomes a user-controllable parameter in several contexts. When generating from a starting image rather than from pure noise (image-to-image generation), the noise level ( often called denoising strength or img2img strength ) controls how much noise is added to the input image before the denoising process begins. A high noise level means more noise is added, giving the model more freedom to diverge from the original image in its reconstruction: essentially a stronger transformation of the input. A low noise level means less noise is added, keeping the model's output closer to the original image: a more conservative transformation. This parameter is fundamental to image-to-image workflows: it determines the balance between preserving the structure of the input image and allowing the model to reinterpret it according to the prompt. At very low noise levels, the output closely mirrors the composition and structure of the input; at very high noise levels, the output uses the input as only a loose reference, with prompt guidance taking precedence.
Noise level also interacts with the step count of the generation process. More denoising steps allow the model to make finer, more incremental adjustments as it removes noise, generally producing higher quality and more detailed outputs: but at the cost of greater compute time. Fewer steps produce faster results that may lack fine detail or exhibit visible artefacts from the coarser denoising resolution. The noise schedule: how rapidly or gradually the model transitions from high noise to low noise across the denoising steps: is another parameter that affects generation quality and style, and is managed by the sampler algorithm selected for the generation. Understanding noise and noise level in this context helps users understand why certain generation behaviours occur: why image-to-image transformations at high strength depart so dramatically from the input, why more steps tend to improve fine detail, and why the same prompt can produce meaningfully different results depending on the random starting noise (the seed).
