Search papers, labs, and topics across Lattice.
The paper introduces ABC, a novel diffusion-based approach for generating continuous-time, continuous-space stochastic processes conditioned on partial observations. ABC addresses limitations of existing diffusion models by modeling the process with a single continual SDE that tracks real time and process states, enabling more stable integration and physically plausible dynamics. By deriving SDE dynamics via changes-of-measure on path space and extending denoising score matching, ABC achieves path-dependent conditioning on arbitrary subsets of the state history and/or future, outperforming existing methods in video generation and weather forecasting.
Forget noisy starting points: ABC leverages the structural similarity between time-adjacent states in continuous processes to generate more realistic videos and weather forecasts.
Generating continuous-time, continuous-space stochastic processes (e.g., videos, weather forecasts) conditioned on partial observations (e.g., first and last frames) is a fundamental challenge. Existing approaches, (e.g., diffusion models), suffer from key limitations: (1) noise-to-data evolution fails to capture structural similarity between states close in physical time and has unstable integration in low-step regimes; (2) random noise injected is insensitive to the physical process's time elapsed, resulting in incorrect dynamics; (3) they overlook conditioning on arbitrary subsets of states (e.g., irregularly sampled timesteps, future observations). We propose ABC: Any-Subset Autoregressive Models via Non-Markovian Diffusion Bridges in Continuous Time and Space. Crucially, we model the process with one continual SDE whose time variable and intermediate states track the real time and process states. This has provable advantages: (1) the starting point for generating future states is the already-close previous state, rather than uninformative noise; (2) random noise injection scales with physical time elapsed, encouraging physically plausible dynamics with similar time-adjacent states. We derive SDE dynamics via changes-of-measure on path space, yielding another advantage: (3) path-dependent conditioning on arbitrary subsets of the state history and/or future. To learn these dynamics, we derive a path- and time-dependent extension of denoising score matching. Our experiments show ABC's superiority to competing methods on multiple domains, including video generation and weather forecasting.
Stanford HAI