Viability of MMD-based drifting fields

Investigate whether a drifting field derived from the Maximum Mean Discrepancy (MMD) objective can yield reasonable ImageNet generation performance within the Drifting Model framework, and identify the kernel, normalization, or algorithmic modifications necessary to make such an MMD-based instantiation work in practice.

Background

The paper relates its drifting-field formulation to MMD by deriving the implicit drift from the gradient of MMD2 and discussing differences such as kernel normalization. Despite the theoretical connection, the authors’ attempts to use an MMD-driven approach did not produce satisfactory results.

Establishing a practical MMD-based realization would bridge classic moment-matching methods with the training-time evolution paradigm and could offer alternative kernels or objectives that retain Drifting Models’ one-step inference benefits.

References

In our experiments, we were not able to obtain reasonable results using the MMD framework.

— Generative Modeling via Drifting  (2602.04770 - Deng et al., 4 Feb 2026) in Appendix — The Drifting Field of MMD