<ul class="dashed" data-apple-notes-indent-amount="0"><li><span style="font-family: '.PingFangUITextSC-Regular'">文章标题:</span>FreeTraj: Tuning-Free Trajectory Control in Video Diffusion Models</li><li><span style="font-family: '.PingFangSC-Regular'">文章地址:</span><a href="https://arxiv.org/abs/2406.16863">https://arxiv.org/abs/2406.16863</a> </li><li>arxiv</li></ul> <img src="https://res.cloudinary.com/montaigne-io/image/upload/v1749031096/4807641D-8EE0-424D-A476-A52B079636EF.png" style="background-color:initial;max-width:min(100%,2370px);max-height:min(956px);;background-image:url(https://res.cloudinary.com/montaigne-io/image/upload/v1749031096/4807641D-8EE0-424D-A476-A52B079636EF.png);height:auto;width:100%;object-fit:cover;background-size:cover;display:block;" width="2370" height="956"> 文章提到FreeInit中揭示的低频信号在加噪过程中保持得比较好,因此探究了初始噪声对生成视频的布局的影响,发现设置相同初始噪声的区域会呈现移动的现象,需要在高频进行重新采样,配合轨迹+高频重采样完成了噪声初始化引导;其次,文章还对不同attention层进行了mask的操作,为了使主体区域关注其轨迹本身以及前景token关注其主体区域以及每帧特定位置的像素关注特定帧的该区域。 <ul class="dashed" data-apple-notes-indent-amount="0"><li>数据:无需训练数据</li><li>指标:FVD;KVD;CLIP-SIM;mIoU;CD</li><li>硬件:不重要</li><li>开源:<a href="http://haonanqiu.com/projects/FreeTraj.html">http://haonanqiu.com/projects/FreeTraj.html</a> </li></ul>