Rolling Diffusion is a novel approach to diffusion models that addresses the limitations of standard diffusion methods in temporal data generation. Unlike traditional diffusion models, which treat all frames equally in terms of noise, Rolling Diffusion uses a sliding window denoising process that assigns more noise to later frames, reflecting greater uncertainty about the future. This approach ensures that the diffusion process progresses through time by increasing noise in later frames, leading to better performance in complex temporal dynamics. The method is evaluated on video prediction tasks using the Kinetics-600 dataset and chaotic fluid dynamics forecasting experiments, where Rolling Diffusion outperforms standard diffusion models. The paper introduces Rolling Diffusion as a framework that explicitly corrupts data from past to future by reparameterizing the global diffusion time to a local time for each frame. This allows for a local sliding window sequential denoising process, which reduces computational complexity and improves performance. Rolling Diffusion is trained with a local time reparameterization that enables the model to focus on frames within a sliding window, leading to more efficient and accurate generation of temporal data. The method is also shown to handle boundary conditions effectively, allowing for end-to-end generation of sequences. Experiments on the Kolmogorov Flow and BAIR Robot Pushing datasets demonstrate that Rolling Diffusion outperforms standard diffusion models in dynamic settings. The model is able to generate high-quality video predictions and handle long temporal rollouts effectively. The results show that Rolling Diffusion consistently outperforms standard diffusion methods, particularly in scenarios with high variability and complexity. The paper concludes that Rolling Diffusion is particularly effective in dynamic settings, where the data is highly variable, and offers exciting future directions in video, audio, and weather or climate modeling.Rolling Diffusion is a novel approach to diffusion models that addresses the limitations of standard diffusion methods in temporal data generation. Unlike traditional diffusion models, which treat all frames equally in terms of noise, Rolling Diffusion uses a sliding window denoising process that assigns more noise to later frames, reflecting greater uncertainty about the future. This approach ensures that the diffusion process progresses through time by increasing noise in later frames, leading to better performance in complex temporal dynamics. The method is evaluated on video prediction tasks using the Kinetics-600 dataset and chaotic fluid dynamics forecasting experiments, where Rolling Diffusion outperforms standard diffusion models. The paper introduces Rolling Diffusion as a framework that explicitly corrupts data from past to future by reparameterizing the global diffusion time to a local time for each frame. This allows for a local sliding window sequential denoising process, which reduces computational complexity and improves performance. Rolling Diffusion is trained with a local time reparameterization that enables the model to focus on frames within a sliding window, leading to more efficient and accurate generation of temporal data. The method is also shown to handle boundary conditions effectively, allowing for end-to-end generation of sequences. Experiments on the Kolmogorov Flow and BAIR Robot Pushing datasets demonstrate that Rolling Diffusion outperforms standard diffusion models in dynamic settings. The model is able to generate high-quality video predictions and handle long temporal rollouts effectively. The results show that Rolling Diffusion consistently outperforms standard diffusion methods, particularly in scenarios with high variability and complexity. The paper concludes that Rolling Diffusion is particularly effective in dynamic settings, where the data is highly variable, and offers exciting future directions in video, audio, and weather or climate modeling.