DriftXpress: Faster Drifting Models via Projected RKHS Fields

2026-05-12

AI Summary

This work addresses the high computational cost of training diffusion models and their limited scalability despite efficient one-step inference. To this end, we propose an acceleration method based on projected reproducing kernel Hilbert space (RKHS) fields. By approximating the diffusion kernel in a low-rank feature space, our approach introduces projected RKHS fields into diffusion modeling for the first time, substantially reducing training costs while preserving the characteristic attraction–repulsion dynamical structure. Experimental results demonstrate that the proposed model achieves FID scores comparable to those of standard diffusion models across multiple image generation benchmarks, yet with significantly reduced training time, thereby reconciling training efficiency with the advantage of fast one-step inference.

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.

Problem

• drifting models
• training cost
• one-step generative modeling
• inference efficiency
• RKHS fields

Innovation

• Drifting Models
• Projected RKHS Fields
• Low-rank Approximation
• One-step Generative Modeling
• Training Efficiency