The noise predictor is trained by adding known noise to images and having it predict that noise

Diffusion training is self-supervised. Take an image from the dataset, generate some noise, and add it — this is one training example. Because the noise amount can be varied from 0 (clean) to total, a single image yields tens of training examples across noise levels. The noise predictor is trained, given a noisy image and a step number, to predict the noise that was added. Once trained, running it in the reverse configuration (repeatedly predicting-and-subtracting noise starting from pure noise) ‘actually creates images’. The model doesn’t reproduce originals; it learns to move a noisy tensor toward the training distribution. The training set therefore determines the model’s style and content (Stable Diffusion v1 used LAION Aesthetics).

Examples

Take a cat photo, add a chosen noise amount, train the UNet to predict that noise. Repeat across many noise levels and millions of images. Amount 0 = no noise, amount 4 = total noise in Alammar’s illustration.

Assessment

Describe the forward-diffusion training loop: what are the noise predictor’s inputs and target during training, and why does learning to predict added noise let the trained model generate images at inference?