研究人员正在探索生成模型的新方法，重点关注Wasserstein梯度流以提高效率和样本质量。一种方法W-Flow，实现了最先进的图像单步生成，与传统的扩散模型相比，采样时间大大缩短。其他论文研究了生成模型的输出优化以及分数-差分流的理论基础，将不同的生成建模技术联系起来，并识别某些流类型的潜在障碍。 AI
影响 Wasserstein梯度流和单步生成方面的进步有望为复杂任务带来更快、更高效的AI模型。
排序理由 多篇arXiv论文详细介绍了生成模型的新理论和算法方法。
AI 生成摘要 · Google Gemini · 来自 7 个来源。 我们如何撰写摘要 →
arXiv:2406.13619v4 Announce Type: replace Abstract: This paper develops a generative model by minimizing the second-order Wasserstein loss (the $W_2$ loss) through a distribution-dependent ordinary differential equation (ODE), whose dynamics involves the Kantorovich potential ass…
arXiv:2509.23800v3 Announce Type: replace Abstract: Modern generative AI models, such as diffusion and flow matching models, can sample from rich data distributions. However, many applications, especially in science and engineering, require more than drawing samples from the mode…
arXiv:2304.12906v4 Announce Type: replace-cross Abstract: Implicit generative modeling (IGM) aims to produce samples of synthetic data matching the characteristics of a target data distribution. Recent work (e.g. score-matching networks, diffusion models) has approached the IGM p…
Diffusion models generate samples by denoising along the score of a perturbed target distribution. In practice, one trains a neural diffusion model, which is computationally expensive. Recent work suggests that score matching implicitly smooths the empirical score, and that this …
arXiv:2605.11755v1 Announce Type: cross Abstract: Diffusion models and flow-based methods have shown impressive generative capability, especially for images, but their sampling is expensive because it requires many iterative updates. We introduce W-Flow, a framework for training …
Diffusion models and flow-based methods have shown impressive generative capability, especially for images, but their sampling is expensive because it requires many iterative updates. We introduce W-Flow, a framework for training a generator that transforms samples from a simple …
arXiv:2604.15439v3 Announce Type: replace Abstract: We study dynamic measure transport for generative modeling, focusing on transport maps that connect a source measure $P_0$ to a target measure $P_1$ by integrating a velocity field of the form $v_t(x) = \mathbb{E}[\dot X_t \mid …