<ul class="dashed" data-apple-notes-indent-amount="0"><li><span style="font-family: '.PingFangUITextSC-Regular'">文章标题:</span>Enhance-A-Video: Better Generated Video for Free</li><li><span style="font-family: '.PingFangSC-Regular'">文章地址:</span><a href="https://arxiv.org/abs/2502.07508">https://arxiv.org/abs/2502.07508</a> </li><li>arxiv</li></ul> <img src="https://res.cloudinary.com/montaigne-io/image/upload/v1752471435/40922C17-72BC-467E-8BF9-85E2DD6A14BA.png" style="background-color:initial;max-width:min(100%,2318px);max-height:min(1206px);;background-image:url(https://res.cloudinary.com/montaigne-io/image/upload/v1752471435/40922C17-72BC-467E-8BF9-85E2DD6A14BA.png);height:auto;width:100%;object-fit:cover;background-size:cover;display:block;" width="2318" height="1206"> 这篇文章中,作者发现在时域的self-attention map中,对角线的权重比其他区域要大,因此作者就提出,在某层的前向过程中,利用attention map评估帧间的权重占比,通过该占比,调整该层的输出与残差的比例(作者通过实验发现某层的输出和残差的比例中,残差占大部分),增加帧间权重大的层的输出比重(通过温度系数调节)。 <a href="../../../../files/Accounts/C037F400-EC11-4FAB-ACA5-467EE47E1BD1/Media/7A918E58-7822-48CF-89E8-773DB1B9CCD1/1_560F839A-F790-4225-B233-0C1AA6BBD7A3/Pasted%20Graphic%206.png" class="attr" data-apple-notes-zidentifier="34D1F3FF-B66E-448B-8C1E-C16142BE3E30"></a> <ul class="dashed" data-apple-notes-indent-amount="0"><li>数据:无需训练</li><li>指标:VBench;User Study</li><li>硬件:未提及</li><li>开源:<a href="https://github.com/NUS-HPC-AI-Lab/Enhance-A-Video">https://github.com/NUS-HPC-AI-Lab/Enhance-A-Video</a> </li></ul>