<?xml version="1.0" encoding="utf-8" standalone="yes" ?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Kai Zhu | ViLab</title>
    <link>https://vilab.team/author/kai-zhu/</link>
      <atom:link href="https://vilab.team/author/kai-zhu/index.xml" rel="self" type="application/rss+xml" />
    <description>Kai Zhu</description>
    <generator>Hugo Blox Builder (https://hugoblox.com)</generator><language>en-us</language><lastBuildDate>Mon, 20 Apr 2026 00:00:00 +0000</lastBuildDate>
    <image>
      <url>https://vilab.team/media/icon_hu2896232876136423579.png</url>
      <title>Kai Zhu</title>
      <link>https://vilab.team/author/kai-zhu/</link>
    </image>
    
    <item>
      <title>Facm: Flow-anchored consistency models</title>
      <link>https://vilab.team/publication/facm-flow-anchored-consistency-models/</link>
      <pubDate>Mon, 20 Apr 2026 00:00:00 +0000</pubDate>
      <guid>https://vilab.team/publication/facm-flow-anchored-consistency-models/</guid>
      <description>&lt;p&gt;本文针对连续时间一致性模型（CM）训练不稳定的问题，指出其根源在于捷径目标导致瞬时速度场被灾难性遗忘。为此提出流锚定一致性模型（FACM），以流匹配任务作为动态锚点，并设计扩展时间间隔策略统一优化、解耦两个任务，实现稳定且架构无关的训练。在ImageNet 256×256上，蒸馏LightningDiT模型取得NFE=2时FID 1.32、NFE=1时FID 1.70的SOTA结果；同时提出内存高效的Chain-JVP，将FACM扩展到140亿参数的Wan 2.2模型，加速文本到图像推理至2-8步。&lt;/p&gt;
</description>
    </item>
    
  </channel>
</rss>
