DriftXpress: Faster Drifting Models via Projected RKHS Fields

TL;DR

DriftXpress introduces a low-rank approximation of drifting kernels in RKHS to accelerate training while maintaining one-step inference quality in generative image models.

Contribution

It presents a novel accelerated drifting model using projected RKHS fields that reduces training cost without sacrificing inference speed or quality.

Findings

DriftXpress achieves comparable FID scores to standard drifting models.

It significantly reduces wall-clock training time.

Maintains one-step inference advantage in image generation.

Abstract

Drifting Models have emerged as a new paradigm for one-step generative modeling, achieving strong image quality without iterative inference. The premise is to replace the iterative denoising process in diffusion models with a single evaluation of a generator. However, this creates a different trade-off: drifting reduces inference cost by moving much of the computation into training. We introduce DriftXpress, an accelerated formulation of drifting models based on projected RKHS fields. DriftXpress approximates the drifting kernel in a low-rank feature space. This preserves the attraction-repulsion structure of the original drifting field while reducing the cost of field evaluation. Across image-generation benchmarks, DriftXpress achieves comparable FID to standard drifting while reducing wall-clock training cost. These results show that the training-inference trade-off of drifting models can be pushed further without giving up their one-step inference advantage.