研究人员开发了对比增强流匹配（CAtFM），一个旨在改进生成模型中内容和风格解耦的新框架。通过将对比正则化集成到可逆流匹配公式中，CAtFM在不需要严格因子化表示的情况下强制执行语义一致性。实验表明，与现有方法相比，CAtFM增强了内容和风格检索，改善了嵌入聚类分离，并提供了对分布偏移的更大鲁棒性。 AI
影响 增强了生成模型在可控内容创建和组合泛化方面的能力。
排序理由 该集群包含一篇详细介绍生成模型新框架的研究论文。
AI 生成摘要 · Google Gemini · 来自 2 个来源。 我们如何撰写摘要 →
完整方法见我们的编辑标准。
arXiv:2607.12404v1 Announce Type: new Abstract: Learning representations that separate content and style is crucial for controllable generation and compositional generalization. However, diffusion and flow-based models trained primarily with generative objectives often produce en…
Learning representations that separate content and style is crucial for controllable generation and compositional generalization. However, diffusion and flow-based models trained primarily with generative objectives often produce entangled or misaligned factors. To address this g…