source: google research: towards demystifying the creativity of diffusion models

level: research

diffusion models are powerful tools for generating images and molecules. they often produce new, plausible data instead of just copying training examples. researchers from google investigated why this happens. they found that creativity is not a random accident but a mathematical result of how neural networks learn. during training, models learn to reverse noise into data. a perfect score function would lead to memorization, pulling every noisy point to an exact training sample. in practice, neural networks learn an approximate, smoothed version of this function.

the smoothing comes from regularization during training, like weight decay. neural networks struggle to learn sharp transitions, so they soften them. in a simple one-dimensional example with two training points, a perfect score function would sharply divide the space, sending all points to one of the two. with smoothing, a gentler slope appears between them. this creates an interpolation zone where points can settle between the training data, generating new values. the stronger the regularization, the smoother the function and the more novel the outputs.

in high-dimensional data like images, the effect is direction-dependent. along directions parallel to the data manifold, smoothing slows down movement, preventing collapse onto training points. perpendicular to the manifold, the score function is already smooth, so movement toward the manifold is not slowed. this lets the model reach the manifold of realistic images while exploring gaps between training examples. the result is a balance: outputs are both realistic and novel. the work suggests that creativity is a predictable outcome of neural network learning, not a mysterious property.

why it matters: understanding this mechanism can help design models that reliably generate novel, useful data for applications like drug discovery and image synthesis, while avoiding simple memorization.

source: google research: towards demystifying the creativity of diffusion models