Influpaint: Using diffusion-based generative models to forecast spatiotemporal influenza

# What Influpaint does

# Why that matters Traditional mechanistic and statistical forecasting approaches can struggle with multimodal uncertainty and unexpected emergent trends. Generative diffusion models learn a rich distribution over whole-season patterns rather than fitting a single trajectory. That lets Influpaint produce multiple diverse, plausible futures, which helps capture complex epidemic shapes across space and time.

# How the method was built and trained The authors trained Influpaint on a hybrid dataset combining real surveillance seasons and simulated epidemic trajectories. Encoding seasons as spatiotemporal images lets the same model handle forecasting and other tasks such as imputing missing data without further training. The team experimented with proportions of surveillance versus simulated data and found the best forecast performance using a mix with about 30% surveillance trajectories and 70% simulated trajectories.

# Evaluation and results

# Practical strengths and features

• Spatiotemporal framing: a single representation handles multiple locations and weeks simultaneously.
• Conditional generation: forecasting is treated as inpainting, which naturally produces probabilistic outputs and multiple sampled trajectories.
• Flexibility: the model can impute missing data and generate complete season realizations without retraining.
• Reproducibility: trained model weights, forecast trajectories, evaluation results, and code are publicly archived (Zenodo and GitHub) for reuse and replication.

# Limitations and behavior noted by the authors The model sometimes produced overconfident projections (notably in 2024–2025). The authors also report that mixing simulated trajectories with surveillance data improved performance, implying sensitivity to the composition of training data and the realism of simulations.

# Where to find the resources Code and training/evaluation artifacts are available in public repositories: the Influpaint codebase on GitHub and archived versions plus model outputs on Zenodo. Real-time FluSight submissions are available in the CDC FluSight forecast repositories under the model name UNC_IDD-InfluPaint.

# Bottom line Influpaint demonstrates that diffusion-based generative models can capture complex spatiotemporal structure in influenza seasons and produce competitive probabilistic forecasts. The approach offers a flexible forecasting framework that can generate diverse futures and perform imputation, but careful calibration and training-data composition matter for reliable uncertainty quantification.