Normalization-equivariant Diffusion Models: Learning Posterior Samplers From Noisy And Partial Measurements

Diffusion models (DMs) have rapidly emerged as a powerful framework for image generation and restoration, achieving remarkable perceptual quality. However, existing DMs are primarily trained in a supervised manner by using a large corpus of clean images. This reliance on clean data poses fundamental challenges in many real-world scenarios, where acquiring noise-free data is hard or infeasible, and only noisy and potentially incomplete measurements are available. While some methods are capable of training DMs using noisy data, they are generally effective only when the amount of noise is very mild or when some additional noise-free data is available. In addition, existing methods for training DMs from incomplete measurements require access to multiple complementary acquisition processes, an assumption that poses a significant practical limitation. Here we introduce the first approach for learning DMs for image restoration using only noisy measurement data from a single operator. As a first key contribution, we show that DMs, and more broadly minimum mean squared error denoisers, exhibit a weak form of scale equivariance linking rescaling in signal amplitude to changes in noise intensity. We then leverage this theoretical insight to develop a denoising score-matching strategy that generalizes robustly to noise levels lower than those present in the training data, thereby enabling the learning of DMs from noisy measurements. To further address the challenges of incomplete and noisy data, we integrate our method with equivariant imaging, a complementary self-supervised learning framework that exploits the inherent invariants of imaging problems, in order to train DMs for image restoration from single-operator measurements that are both incomplete and noisy. We validate the effectiveness of our approach through extensive experiments on image denoising, demosaicing, and inpainting, along with comparisons with the state of the art.