JIVE：基于雅可比信息的体积扩展用于多样化生成采样
JIVE: Jacobian-Informed Volume Expansion for Diverse Generative Sampling
浏览论文内容
中文总结 AI 辅助
针对生成模型模式坍缩和样本多样性不足的问题，提出无需训练的JIVE框架，利用端点雅可比矩阵的奇异子空间注入速度扰动，在保持质量的同时最大化端点多样性，并在少步和单步生成中提升像素与特征级多样性。
中文摘要 AI 辅助
生成模型常常面临模式坍缩和样本多样性有限的问题。尽管先前的工作尝试通过联合生成一批样本并排斥其轨迹来缓解这一问题，但这些启发式方法并未显式地最大化最终端点的多样性。我们提出了JIVE，一个无需训练的框架，通过注入与生成器端点雅可比矩阵的前导右奇异子空间对齐的速度扰动来增强生成多样性。通过利用这种局部几何结构，JIVE在保持样本质量的同时，可证明地最大化端点多样性。为了保持实际效率，我们通过基于经典数值线性代数的无矩阵迭代来计算这些扰动方向，仅需少量计算开销。在不同基准测试中，JIVE在少步和单步生成中均提升了像素级和特征级多样性。
英文摘要
Generative models often suffer from mode collapse and limited sample diversity. While prior works attempt to mitigate this by jointly generating a batch of samples and repelling their trajectories, these heuristics do not explicitly maximize the diversity of the resulting endpoints. We introduce JIVE, a training-free framework that enhances generative diversity by injecting velocity perturbations aligned with the leading right singular subspace of the generator's endpoint Jacobian. By leveraging this local geometric structure, JIVE provably maximizes endpoint diversity while preserving sample quality. To maintain practical efficiency, we compute these perturbation directions via matrix-free iterations rooted in classical numerical linear algebra, requiring only a small computational overhead. Across different benchmarks, JIVE boosts both pixel and feature-level diversity in few-step and one-step generation.
发表机构
- New York University（纽约大学）
- Stony Brook University（石溪大学）
- Rutgers University（罗格斯大学）
机构由 AI 辅助整理，请以论文原文为准。
