BAAI says standard continuous diffusion schedules waste over 90% of training budget

BAAI says standard continuous diffusion noising trajectories waste over 90% of the optimization budget.

According to BAAI, tokens keep their original identity for most of the trajectory, then collapse abruptly in a narrow window near the end. Previous schedules spread training steps uniformly, so the model trains on stages where token identity does not change.

The claim comes from a BAAI post on X dated 2026-11-03. The post reports no experiments or results for an alternative schedule.