A diffusion model steers noise toward the training distribution, not toward exact stored images

A key conceptual point: subtracting the predicted noise moves an image toward the training data’s distribution, not toward any exact stored image — ‘the world of pixel arrangements where the sky is usually blue and above the ground, people have two eyes, cats have pointy ears’. The model learns these statistical regularities rather than memorizing samples, so it can produce novel images that fit the same patterns, generalize to unseen combinations, and be steered by prompts toward different regions of the distribution. It also means outputs inherit the training data’s aesthetics and biases: train on aesthetic images (LAION Aesthetics) and outputs look aesthetic; train on logos and it makes logos.

Examples

A model trained on LAION Aesthetics tends to produce aesthetically pleasing images; one trained on logos produces logos. Unusual prompts yield plausible hybrids because the model interpolates within the learned distribution.

Assessment

Explain why a diffusion model can generate ‘a purple elephant in a cubist style’ even if no such image was in training, and what it means to say the model learned a distribution rather than memorized images.