Training architecture

A denoising diffusion model learns a distribution over season images. Training adds Gaussian noise at a diffusion time and teaches a U-Net to predict it. Sampling reverses this process from noise to a generated season.

The selected paper model uses 500 diffusion steps with a cosine schedule and a ResNet U-Net with channel multipliers (1,2,2,4). Each input has one channel and a 64 × 64 week/location grid. The model uses the July 17 30S70M dataset, square-root preprocessing, and no additional training enrichment.

Dataset preparation mixes surveillance and simulated sources and completes frames before training. Training-time preprocessing is defined by the selected configuration in influpaint/batch/config.py; the weights and their corresponding dataset must be retained together.

See step 2 for the image construction and step 3 for the paper candidates, Slurm command, and archived checkpoint.