The work analyzes how a denoiser’s output changes under small input perturbations by examining the eigenvalues and eigenvectors of its Jacobian. Across pretrained SiT models, better generative metrics were associated with larger principal Jacobian eigenvalues.
It then introduces input perturbations during training that amplify responses along principal directions or suppress noisy, data-irrelevant responses without explicitly computing an eigendecomposition. ImageNet experiments indicate that both objectives improve generative metrics and produce cumulative gains when combined, making Jacobian spectra useful both for diagnosing denoisers and designing training objectives.