用低秩特征空间加速生成模型，一步完成图像生成且训练更省时。
DriftXpress: Faster Drifting Models via Projected RKHS Fields
- 基于投影再生核希尔伯特空间，用低秩近似优化生成场计算。
- 在多个图像生成基准上达到与原方法相当的生成质量（FID），训练耗时显著降低。
- 适合追求高效训练与快速生成的科研与工程应用者。
漂移模型（Drifting Models）作为一种新型单步生成范式，无需迭代去噪即可实现高质量图像生成。其核心是将扩散模型中的迭代去噪过程替换为生成器的一次性评估，但代价是将大量计算转移至训练阶段。本文提出 DriftXpress，一种基于投影再生核希尔伯特空间（Projected RKHS）场的加速漂移模型形式。该方法在低秩特征空间中近似漂移核，保留原始漂移场的吸引-排斥结构，同时降低场评估开销。在多个图像生成基准上，DriftXpress 实现了与标准漂移模型相当的 FID 性能，同时大幅减少实际训练时间。结果表明，漂移模型的训练-推理权衡可进一步优化，且不牺牲其单步推理优势。
原文摘要 · Abstract (English)
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.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。