https://vilab.team/tag/%E5%B0%91%E6%AD%A5%E7%94%9F%E6%88%90/ 少步生成 Hugo Blox Builder (https://hugoblox.com)en-usMon, 20 Apr 2026 00:00:00 +0000 https://vilab.team/media/icon_hu2896232876136423579.png https://vilab.team/tag/%E5%B0%91%E6%AD%A5%E7%94%9F%E6%88%90/ https://vilab.team/publication/facm-flow-anchored-consistency-models/ Mon, 20 Apr 2026 00:00:00 +0000 https://vilab.team/publication/facm-flow-anchored-consistency-models/ <p>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。</p> https://vilab.team/event/%E7%A5%9D%E8%B4%BA%E5%AE%9E%E9%AA%8C%E5%AE%A4%E7%A7%91%E7%A0%94%E6%88%90%E6%9E%9C%E5%8F%91%E8%A1%A8%E4%BA%8E-iclr-2026/ Mon, 20 Apr 2026 00:00:00 +0000 https://vilab.team/event/%E7%A5%9D%E8%B4%BA%E5%AE%9E%E9%AA%8C%E5%AE%A4%E7%A7%91%E7%A0%94%E6%88%90%E6%9E%9C%E5%8F%91%E8%A1%A8%E4%BA%8E-iclr-2026/ <p>热烈祝贺彭岩松同学！论文《Facm: Flow-anchored consistency models》已发表在 <em>ICLR</em> 2026。</p> <h2 id="facm-flow-anchored-consistency-models">Facm: Flow-anchored consistency models</h2> <p><strong>祝贺彭岩松同学！该论文已发表在 <em>ICLR</em> 2026。</strong></p> <p> <figure > <div class="d-flex justify-content-center"> <div class="w-100" ><img alt="Facm: Flow-anchored consistency models" srcset=" /event/%E7%A5%9D%E8%B4%BA%E5%AE%9E%E9%AA%8C%E5%AE%A4%E7%A7%91%E7%A0%94%E6%88%90%E6%9E%9C%E5%8F%91%E8%A1%A8%E4%BA%8E-iclr-2026/images/paper-01_hu12002391219066891609.webp 400w, /event/%E7%A5%9D%E8%B4%BA%E5%AE%9E%E9%AA%8C%E5%AE%A4%E7%A7%91%E7%A0%94%E6%88%90%E6%9E%9C%E5%8F%91%E8%A1%A8%E4%BA%8E-iclr-2026/images/paper-01_hu15036098972298221130.webp 760w, /event/%E7%A5%9D%E8%B4%BA%E5%AE%9E%E9%AA%8C%E5%AE%A4%E7%A7%91%E7%A0%94%E6%88%90%E6%9E%9C%E5%8F%91%E8%A1%A8%E4%BA%8E-iclr-2026/images/paper-01_hu3776373299308260365.webp 1200w" src="https://vilab.team/event/%E7%A5%9D%E8%B4%BA%E5%AE%9E%E9%AA%8C%E5%AE%A4%E7%A7%91%E7%A0%94%E6%88%90%E6%9E%9C%E5%8F%91%E8%A1%A8%E4%BA%8E-iclr-2026/images/paper-01_hu12002391219066891609.webp" width="760" height="207" loading="lazy" data-zoomable /></div> </div></figure> </p> <ul> <li><strong>作者：</strong> Yansong Peng、Kai Zhu、Yu Liu、Pingyu Wu、Hebei Li、Xiaoyan Sun、Feng Wu</li> <li><strong>发表载体：</strong> <em>ICLR</em> 2026</li> <li><strong>发表时间：</strong> 2026年4月20日</li> <li><strong>相关链接：</strong> <a href="https://proceedings.iclr.cc/paper_files/paper/2026/hash/0d0dac08f4199f0c348dd2feace0305a-Abstract-Conference.html" target="_blank" rel="noopener">论文链接</a> · <a href="https://proceedings.iclr.cc/paper_files/paper/2026/file/0d0dac08f4199f0c348dd2feace0305a-Paper-Conference.pdf" target="_blank" rel="noopener">PDF</a> · <a href="https://github.com/ali-vilab/FACM" target="_blank" rel="noopener">代码</a></li> </ul> <h3 id="论文介绍">论文介绍</h3> <p>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。</p>