Denoising Strain: A Geometric Order Parameter That Reveals Learning Ahead of Eval Metrics

Diffusion models are evaluated through denoising loss and sample-quality metrics such as Fréchet Inception Distance (FID). These scores track performance, but do not directly show the generative structure taking shape during learning. We introduce Denoising Strain, a geometric description of how each complete sampler step contracts, amplifies, and transports nearby perturbations. Separating overall scale from directional stretching gives a tensor order parameter: it vanishes when all directions are treated equally and develops structure as preferred directions form. The discovered axes have measurable consequences for generation. On pretrained animal-face and human-face models, leading-axis interventions produce median final-image separations about 28 and 24 times those of equal-norm random controls. A separate continuous-step study uses three intervention doses on 100 held-out trajectories per dataset and finds feature-space dose-response slopes above controls at all 11 measured states. We also track this geometry along uninterrupted MNIST and CIFAR-10 training runs, where initially near-equal stretches differentiate and expanding directions appear. On CIFAR-10, expansion and anisotropy are detected after 10,000–20,000 training images, whereas the checkpoint completing half of the total log-FID improvement arrives after 5 million images—250–500 times later in training exposure. Denoising Strain thus provides an order parameter that links how diffusion models generate to how that ability develops during training.