DIMOND：DIffusion Model OptimizatioN with Deep learning

The accurate estimation of diffusion model parameter values using non-linear optimization is time-consuming. Supervised learning methods using neural networks (NNs) are faster and more accurate but require external ground-truth data for training. A unified and self-supervised learning-based diffusion model estimation method DIMOND is proposed. DIMOND maps diffusion data to model parameter values using NNs, synthesizes the input data from the predictions using the forward model, and minimizes the difference between the raw and synthetic data. DIMOND outperforms conventional ordinary least square regression (OLS) and has a high potential to improve and accelerate data fitting for more complicated diffusion models.