Extreme floods are becoming more frequent and destructive due to climate change, yet predicting them remains difficult. Existing models often underestimate peak river flows and fail to effectively quantify uncertainty due to the scarcity of peak flow samples in historical datasets. In this study, we introduce DRUM, a new deep learning approach based on diffusion models, a cutting-edge generative AI technique. Tested in representative U.S. river basins, DRUM outperforms state-of-the-art deep learning models for the most extreme floods in over 70% of cases and extends early warnings by nearly a full day for major floods. When given actual rainfall data instead of weather forecasts, DRUM’s accuracy further improves, increasing recall by 0.3–0.4 and extending early warnings by 2.3 days for the most extreme floods. In regions where floods are mainly driven by heavy rainfall, such as the eastern and northwestern United States, it further extends warning times by up to 7 days. These results demonstrate how AI can enhance flood prediction and disaster preparedness while highlighting the importance of accurate rainfall forecasts. Beyond flood forecasting, this study illustrates the broader potential of generative AI in improving predictions of extreme environmental events and advancing Earth system science.