The Deeper Mathematics of Generative Modeling via Drifting

By Nripendra Acharya

The paradigm of drifting models represents the cutting edge of generative AI research. It belongs to the family of AI systems that produce realistic synthetic data, of which image generators are the most prominent example.

This technique was introduced in February 2026 by researchers at MIT and Harvard, including Kaiming He, one of the most cited figures in deep learning. The work is recent and still maturing, but it warrants attention from startups and companies for a specific reason: it generates a sample in a single forward pass of the network, where the diffusion models underlying tools such as Midjourney and DALL·E must run tens of iterative refinement steps per sample.

Holding quality constant, this reduces inference compute — saving cost. Currently, the method achieves state-of-the-art results for one-step generation on standard image benchmarks.

The relevant paper is Generative Modeling via Drifting by Deng et al.

To explain the breakthrough in terms of its deeper mathematics, we can describe the following equation intuitively and succinctly.

Notice, in the brackets, the key detail: there is bilinear gating between two kernels — (1) the sample distribution with itself, a repulsion distance, and (2) the sample distribution with the data distribution, an attraction distance.

This gating is applied against (3) the vector difference between attraction and repulsion — the y terms — and is normalized, producing the drifting field.

The drift field then defines the training target.

The deep mathematical intuition: interpreted in terms of repulsion and attraction and the bilinear gating, the drifting field combined with the frozen target gives each iteration a curl toward the data distribution.

Put simply, the sample feels a pull toward real data and a push away from other generated data. The network is repeatedly updated so that the overall trajectory of the curl converges on the real data distribution.