<ul class="dashed" data-apple-notes-indent-amount="0"><li><span style="font-family: '.PingFangUITextSC-Regular'">文章标题:</span>FASTERCACHE: TRAINING-FREE VIDEO DIFFUSION MODEL ACCELERATION WITH HIGH QUALITY</li><li><span style="font-family: '.PingFangSC-Regular'">文章地址:</span><a href="https://arxiv.org/abs/2410.19355">https://arxiv.org/abs/2410.19355</a> </li><li>ICLR 2025</li></ul> <img src="https://imagedelivery.net/phxEHgsq3j8gSnfNAJVJSQ/node3_0ac3f819-577a-4963-8337-7ba24b18356c/public" style="background-color:initial;max-width:min(100%,2474px);max-height:min(946px);;background-image:url(https://imagedelivery.net/phxEHgsq3j8gSnfNAJVJSQ/node3_0ac3f819-577a-4963-8337-7ba24b18356c/public);height:auto;width:100%;object-fit:cover;background-size:cover;display:block;" width="2474" height="946"> 这篇文章的主要出发点为减少CFG中uncondition部分的计算,作者通过对采样过程中cond和uncond在同一步骤/相邻步骤的差进行可视化,发现其具有U型的形状,如下图: <img src="https://imagedelivery.net/phxEHgsq3j8gSnfNAJVJSQ/node3_202f5d14-6a94-450c-a3f1-8f2cbc5081d9/public" style="background-color:initial;max-width:min(100%,2072px);max-height:min(876px);;background-image:url(https://imagedelivery.net/phxEHgsq3j8gSnfNAJVJSQ/node3_202f5d14-6a94-450c-a3f1-8f2cbc5081d9/public);height:auto;width:100%;object-fit:cover;background-size:cover;display:block;" width="2072" height="876"> 其中,同一步数下cond和uncond的差值非常小。因此作者提出了CFG-Cache。首先作者提到,直接用cond替换uncond的效果不好,随后分析了两者在高频和低频上的差异,如下图: <img src="https://imagedelivery.net/phxEHgsq3j8gSnfNAJVJSQ/node3_c805856e-f1af-45b4-96f3-3b0667a03c02/public" style="background-color:initial;max-width:min(100%,2050px);max-height:min(614px);;background-image:url(https://imagedelivery.net/phxEHgsq3j8gSnfNAJVJSQ/node3_c805856e-f1af-45b4-96f3-3b0667a03c02/public);height:auto;width:100%;object-fit:cover;background-size:cover;display:block;" width="2050" height="614"> 随后提出用高频和低频信息来修正cond,对uncond做一个更加精准的估计。 <img src="https://imagedelivery.net/phxEHgsq3j8gSnfNAJVJSQ/node3_ebb363b1-a954-4db6-9373-6cf0e619c614/public" style="background-color:initial;max-width:min(100%,1674px);max-height:min(156px);;background-image:url(https://imagedelivery.net/phxEHgsq3j8gSnfNAJVJSQ/node3_ebb363b1-a954-4db6-9373-6cf0e619c614/public);height:auto;width:100%;object-fit:cover;background-size:cover;display:block;" width="1674" height="156"> <img src="https://imagedelivery.net/phxEHgsq3j8gSnfNAJVJSQ/node3_4c8779b8-c159-43c7-82a3-98b264282ba7/public" style="background-color:initial;max-width:min(100%,1622px);max-height:min(194px);;background-image:url(https://imagedelivery.net/phxEHgsq3j8gSnfNAJVJSQ/node3_4c8779b8-c159-43c7-82a3-98b264282ba7/public);height:auto;width:100%;object-fit:cover;background-size:cover;display:block;" width="1622" height="194"> <ul class="dashed" data-apple-notes-indent-amount="0"><li>开源:<a href="https://github.com/Vchitect/FasterCache">https://github.com/Vchitect/FasterCache</a> </li></ul>