InvDiff

Summary

InvDiff is an invariant-guidance method for mitigating unknown biases in pretrained text-to-image diffusion models without auxiliary bias labels. It also includes a limited AusElec/TimeGrad OOD forecasting experiment.

Official Artifacts

• Paper: https://arxiv.org/abs/2412.08480
• Code: https://github.com/Hundredl/InvDiff

Role In The Wiki

InvDiff is primarily a text-to-image diffusion debiasing method. It is adjacent to time-series generation and has limited OOD forecasting evidence, so use it as a shortcut and unknown-bias mitigation pattern rather than broad TSFM generation evidence.

Related Pages

• Time-Series Generation
• Distribution Priors In Self-Supervised Learning