Drifting Models

Generative Modeling via Drifting

Mingyang Deng, He Li, Tianhong Li, Yilun Du, Kaiming He · arXiv 2026

Read paper · arXiv:2602.04770

Construct a drifting field that attracts real samples and repels generated samples, then train a generator against a stopped-gradient target.

What you will implement

Implement the drifting field and stopped-gradient loss, and verify generator behavior on a small two-dimensional distribution.

• 18. Drifting: Pairwise Field and Frozen Targets

  Implement the distance-kernel drifting field and its stop-gradient target; train a generator with one-pass inference.

  Implement drift_field, drifting_loss. toy.py provides the training and visualization scaffold.