Noise is a random or unreliable component of an image or dataset that obscures meaningful signal. It includes invalid inputs, mislabeled examples, ambiguous labels, and accidental correlations. In image generation, pure noise becomes the starting point for repeatedly reconstructing a plausible image. section 5.1.1 section 17.2
Adding noise to MNIST
An MNIST experiment appends a random noise channel to every flattened image, then trains the same classifier on versions with noise channels and all-zero channels. Comparing validation accuracy reveals whether the model is relying on irrelevant input variation.
Noise appears in several forms:
- Invalid or unusual inputs, such as an all-black MNIST digit. section 5.1.1
- Incorrect labels, which can make nearby valid examples receive the wrong prediction. section 5.1.1
- Ambiguity and randomness, where identical measurements can support different outcomes. section 5.1.1
- Synthetic noise, which denoising models gradually remove to recover or hallucinate image structure. section 17.2
Diffusion models treat denoising as an iterative process: begin with pure noise, remove a small amount at each step, and finish with an image dominated by signal. Diffusion time tracks progress, while a schedule controls the changing noise and signal rates. section 17.2 section 17.2.3
