# BAAI says standard diffusion schedules waste most of their training budget

> Original title: Why did previous continuous diffusion struggle?

- Source: BAAI
- Published: 2026-11-03T14:03:01.000Z
- Bad Influence: https://badinfluence.news/items/bwakdnljldlo2y22dl4z8synv
- Original: https://x.com/BAAIBeijing/status/2117618553247445122

## Summary

BAAI (@BAAIBeijing) says standard continuous diffusion trajectories waste over 90% of the optimization budget. Tokens keep their identity for most of the noising trajectory, then collapse abruptly in a narrow window near the end. Previous schedules spread training steps uniformly, so they train the model where token identity does not change.

## Post

Why did previous continuous diffusion struggle?

Standard trajectories waste over 90% of the optimization budget. Tokens retain their original identity for most of the noising trajectory, then collapse abruptly in a tiny window near the end!

Previous schedules allocated training steps uniformly, training the model where token identity didn't even change.

[Video](https://badinfluence.news/api/img-proxy?u=https%3A%2F%2Fpbs.twimg.com%2Ftweet_video_thumb%2FHWKyYayaYAA2PBw.jpg&mode=thumb&exp=1793923200&sig=83f2425d7a7b4154)