<p>Methods behind Influpaint.</p>

<p><b>a.</b> Epidemic seasons are represented as images with time and space as axes, and pixel intensity corresponding to incident hospitalizations. <b>b.</b> Denoising Diffusion Probabilistic Models generate data by training a U-Net to restore an image that has been progressively corrupted by Gaussian noise. <b>c.</b> RePaint algorithm for forecasting. The ground truth and a mask (white: values to preserve; green: values to discard and replace by the generation process) are provided.

At each denoising step, we stitch the noised ground truth and the intermediate denoised sample to obtain the updated trajectory.</p>