<rss xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title>Papers - 分类 - mywebsite</title><link>https://steven-yl.github.io/mywebsite/categories/papers/</link><description>Papers - 分类 - mywebsite</description><generator>Hugo -- gohugo.io</generator><language>zh-CN</language><managingEditor>steven@gmail.com (Steven)</managingEditor><webMaster>steven@gmail.com (Steven)</webMaster><copyright>This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.</copyright><lastBuildDate>Wed, 25 Mar 2026 00:00:00 +0800</lastBuildDate><atom:link href="https://steven-yl.github.io/mywebsite/categories/papers/" rel="self" type="application/rss+xml"/><item><title>Generative Modeling via Drifting</title><link>https://steven-yl.github.io/mywebsite/generative-modeling-via-drifting/</link><pubDate>Wed, 04 Mar 2026 00:00:00 +0000</pubDate><author><name>Steven</name><uri>https://github.com/steven-yl</uri></author><guid>https://steven-yl.github.io/mywebsite/generative-modeling-via-drifting/</guid><description><![CDATA[<h2 id="drifting-models-算法原理详细解析含完整公式" class="headerLink">
    <a href="#drifting-models-%e7%ae%97%e6%b3%95%e5%8e%9f%e7%90%86%e8%af%a6%e7%bb%86%e8%a7%a3%e6%9e%90%e5%90%ab%e5%ae%8c%e6%95%b4%e5%85%ac%e5%bc%8f" class="header-mark"></a>Drifting Models 算法原理详细解析（含完整公式）</h2><p>Drifting Models 是一种面向生成式建模的全新范式，核心创新在于将“分布推送”过程从推理阶段转移到训练阶段，通过引入<strong>漂移场（Drifting Field）</strong>  govern 样本分布的演化，最终实现<strong>单步推理（1-NFE）</strong> 下的高质量生成。该算法突破了扩散模型、流匹配等传统方法依赖多步迭代的效率瓶颈，在 ImageNet 256×256 生成任务中达到当前单步方法的最优性能（ latent 空间 FID 1.54，像素空间 FID 1.61），同时可扩展至机器人控制等其他领域。以下从核心思想、算法架构、关键模块、完整公式及实验验证等方面逐层解析。</p>]]></description></item><item><title>Minimal-Drifting-Models</title><link>https://steven-yl.github.io/mywebsite/minimal-drifting-models-%E7%AE%97%E6%B3%95%E4%B8%8E%E5%AE%9E%E7%8E%B0%E8%AF%B4%E6%98%8E/</link><pubDate>Tue, 03 Mar 2026 12:00:00 +0800</pubDate><author><name>Steven</name><uri>https://github.com/steven-yl</uri></author><guid>https://steven-yl.github.io/mywebsite/minimal-drifting-models-%E7%AE%97%E6%B3%95%E4%B8%8E%E5%AE%9E%E7%8E%B0%E8%AF%B4%E6%98%8E/</guid><description>基于 &lt;a href="https://github.com/Algomancer/Minimal-Drifting-Models" target="_blank" rel="noopener noreferrer">Algomancer/Minimal-Drifting-Models&lt;/a> 与论文 Generative Modeling via Drifting (Deng et al., 2026)，详解漂移场 V、训练损失、推理 1-NFE，以及 Sinkhorn / 特征函数 / 自编码器扩展。</description></item><item><title>DiffusionDriveV2 网络结构图</title><link>https://steven-yl.github.io/mywebsite/diffusiondrivev2_network/</link><pubDate>Wed, 25 Mar 2026 00:00:00 +0800</pubDate><author><name>Steven</name><uri>https://github.com/steven-yl</uri></author><guid>https://steven-yl.github.io/mywebsite/diffusiondrivev2_network/</guid><description>本文以结构图的方式梳理 DiffusionDrive V2 的关键模块与连接关系：双流骨干特征融合、Transformer Decoder×3、扩散式 TrajectoryHead 的截断生成以及粗筛-精筛评分流水线。</description></item><item><title>DiffusionDriveV2 代码结构图</title><link>https://steven-yl.github.io/mywebsite/diffusiondrivev2_code_structure/</link><pubDate>Wed, 25 Mar 2026 00:00:00 +0800</pubDate><author><name>Steven</name><uri>https://github.com/steven-yl</uri></author><guid>https://steven-yl.github.io/mywebsite/diffusiondrivev2_code_structure/</guid><description>通过目录树与模块依赖关系，梳理 DiffusionDrive v2（RL/Selection 相关模型及扩散模块）在工程中的位置与调用链路。</description></item></channel></rss>