Predicting floods accurately is crucial as climate change leads to more frequent and severe flooding events worldwide. Traditional forecasting methods often simplify how floods happen, making assumptions that may not fully capture reality. To address this, we developed a new type of diffusion-based generative model known as the Denoising Diffusion Probabilistic Model (DDPM)—that learns distribution of discharges directly from historical basin discharge data and generates multiple realistic scenarios ensembles, providing better predictions and clearer estimates of uncertainty. We tested our approach globally, comparing it to a state-of-the-art forecasting method (LSTM) across nearly 4,000 basins under various climates. The results showed that our model consistently provided more accurate flood forecasts, especially in areas with unpredictable flooding patterns like hot and dry regions. It also predicted extreme floods better, which are particularly important for risk management. Although human activities, such as reservoir operations, still pose challenges, our model handles these disruptions more effectively than existing methods. Our findings highlight the potential of this approach to significantly improve flood forecasting and flood risk management strategies worldwide.