DriftXpress: Faster Drifting Models via Projected RKHS Fields
DriftXpress is a new formulation of drifting models for one-step generative modeling, introduced in a paper on arXiv. It accelerates training by approximating the drifting kernel in a low-rank feature space using projected RKHS fields, preserving the attraction-repulsion structure while reducing field evaluation cost. The method achieves comparable FID scores to standard drifting on image-generation benchmarks while cutting wall-clock training time. This pushes the training-inference trade-off further, making drifting models more efficient without iterative inference.
Key facts
- DriftXpress is an accelerated formulation of drifting models.
- It uses projected RKHS fields to approximate the drifting kernel.
- The method reduces the cost of field evaluation.
- It preserves the attraction-repulsion structure of the original drifting field.
- DriftXpress achieves comparable FID to standard drifting.
- It reduces wall-clock training cost on image-generation benchmarks.
- The paper is available on arXiv with ID 2605.12183.
- Drifting models replace iterative denoising with a single generator evaluation.
Entities
Institutions
- arXiv