Diff-MN

Summary

Diff-MN is a continuous time-series generation framework for irregular observations. It uses MoE-NCDE dynamics and diffusion-parameterized sample-specific dynamics weights.

Official Artifacts

• Paper: https://arxiv.org/abs/2601.13534
• Code: https://github.com/microsoft/TimeCraft/tree/main/Diff-MN

Role In The Wiki

Diff-MN is the TimeCraft branch most directly connected to irregular and continuous-time latent-state modeling.

Related Pages

• TimeCraft
• Time-Series Generation
• Latent-State Time-Series Modeling
• Mixture Of Experts