Generative Modeling via Drifting

TL;DR

New paradigm: Drifting Models evolve the model's pushforward distribution during training (instead of iterating at inference time like diffusion/flow).
Drifting field: We introduce a drifting field V that moves samples; by design, V vanishes when the generated distribution matches data (equilibrium).
Algorithm: Estimate V from mini-batches, create a "drifted target", and regress the network to match it. Inference is a single forward pass.
Flexibility: Works with representation-space loss and classifier-free-guidance.
Strong performance: ImageNet 256×256: State-of-the-art 1-NFE FID 1.54 (latent) and 1.61 (pixel).

Abstract

Generative modeling can be formulated as learning a mapping f such that its pushforward distribution matches the data distribution. The pushforward behavior can be carried out iteratively at inference time, e.g., in diffusion/flow-based models. In this paper, we propose a new paradigm called Drifting Models, which evolve the pushforward distribution during training and naturally admit one-step inference. We introduce a drifting field that governs the sample movement and achieves equilibrium when the distributions match. This leads to a training objective that allows the neural network optimizer to evolve the distribution. In experiments, our one-step generator achieves state-of-the-art results on ImageNet 256×256, with FID 1.54 in latent space and 1.61 in pixel space.

Resources

Drifting model demo

Interactive notebook demonstrating the core algorithm on toy 2D distributions.

Open in Google Colab

Notebook preview

Release

ImageNet training code and models will be released.

Citation

@article{deng2026drifting,
  title={Generative Modeling via Drifting},
  author={Deng, Mingyang and Li, He and Li, Tianhong and Du, Yilun and He, Kaiming},
  journal={arXiv preprint arXiv:2602.04770},
  year={2026}
}

Generative Modeling via Drifting

Training dynamics

TL;DR

Abstract

Generated Samples

Resources

Citation