We introduce a novel framework, DYffusion, for large-scale probabilistic forecasting. It harnesses the power of diffusion models while fundamentally reimagining their temporal dynamics. By directly coupling the temporal dynamics of the data with the diffusion process, our approach overcomes key limitations of traditional diffusion models. DYffusion achieves significantly faster sampling speeds and reduced memory requirements while generating stable and accurate probabilistic forecasts over extended time horizons. We showcase its effectiveness through an application to climate model emulation, where it successfully generates fast and accurate 10- to 100-year-long global climate simulations at 6-hourly resolution.

Despite advances in diffusion-based text-to-music (TTM) methods, efficient, high-quality generation remains a challenge. We introduce Presto!, an approach to inference acceleration for score-based diffusion transformers via reducing both sampling steps and cost per step. To reduce steps, we develop a new score-based distribution matching distillation (DMD) method for the EDM-family of diffusion models, the first GAN-based distillation method for TTM. To reduce the cost per step, we develop a simple, but powerful improvement to a recent layer distillation method that improves learning via preserving hidden state variance. Finally, we combine our improved step and layer distillation methods together for a dual-faceted approach. We evaluate our step and layer distillation methods independently and show each yield best-in-class performance. Furthermore, we find our combined distillation method can generate high-quality outputs with improved diversity accelerating our base model by 10-18x (32 second output in 230ms, 15x faster than the comparable SOTA model) -- the fastest high-quality TTM model to our knowledge.